|Allegro CL version 10.1|
Unrevised from 10.0 to 10.1.
This document contains the following sections:1.0 Garbage collection introduction
Lisp does its own memory management. The garbage collector is part of the memory management system. It disposes of objects that are no longer needed, freeing up the space that they occupied. While the garbage collector is working, no other work can be done. Therefore, we have made the garbage collector as fast and unobtrusive as possible. Allegro CL uses a version of the generation-scavenging method of garbage collection. Because optimal performance of generation-scavenging garbage collection depends on the application, you have a great deal of control over how the garbage collector works. In this section, we will describe the user interface to the garbage collector, and suggest how to tune its performance for an application. In what follows, the generation-scavenging garbage collection system will be abbreviated gsgc, and the act of garbage collecting will be abbreviated gc.
Symbols naming gc-related functionality are generally in
:excl package or the
The Allegro CL garbage collector is a two-space, generation-scavenging system. The two spaces are called newspace and oldspace. Note that, as we describe below, newspace is divided into two pieces, called areas, and oldspace may be divided into a number of pieces, also called areas. Generally, when we say newspace, we mean both newspace areas and when we say oldspace, we mean all oldspace areas. We try to use the word area when we want to refer to a single area, but please note that this naming convention is new and you may run into text that uses `oldspace' to refer to an oldspace area. Usually, the context should make this clear.
Newspace is gc'ed with operations called scavenges. Scavenges eliminate all garbage in newspace and may retain live objects in newspace or move them to oldspace. Scavenges are typically very fast and reasonably frequent. They are not under user control, but rather triggered automatically when necessary. The section Section 1.5 How newspace is managed with scavenges below describes scavenges.
Oldspace is gc'ed with operations called global gc's. Global gc's are under user control, though triggering them can be automated. Global gc's usually take much longer than scavenges. Oldspaces and global gc's are described throughout this document, but the section Section 1.8 Objects that survive long enough are tenured to oldspace provide an introduction. Global gc's can use multiple (concurrent in SMP Lisps) threads, as we describe in the section Section 1.2 Using multiple threads in global gc's below.
The garbage collector can use multiple threads for global gc's. This is true even in non-SMP Lisps (except, as noted, on Solaris platforms). When extra cores are available, this can have a significant effect on global gc time, since the gc worker threads can run concurrently.
The function sys:gc-parameters returns (or prints) information about aspects of the gc setup. The related sys:gc-parameter returns information about specific parameters and sys:gc-switch returns information about gc switches. Where appropriate, they can be used with setf to reset values.
The two parameters relevant to multithread global gc's are:
:helper-thread-limit: the maximum number of threads that could be used for gc work.
:helper-threads-requested: the number of worker threads desired. This value is setf'able but cannot be larger than the
There is always one gc thread, so helper threads are in addition to
that thread. Setting
0 means that gc's will always use 1 thread only. Users might vary
:helper-threads-requested to see what effect it
has on pause times. There is no reason to
:helper-threads-requested larger than the
number of cores available minus 1. The number of cores available is
the actual number of cores less any that are dedicated to other
purposes, such as managing video.
See Section 5.5 Parameters that use of multiple threads by the GC for more
Some Lisp values exist on the stack. The garbage collector must scan these values to determine whether they are pointing to values in the heap which are, therefore, live and should not be gc'ed.
However, some values on the stack may not be real. When a variable is placed on the stack or when the stack pointer is moved so the stack is larger, values are not initialized (since doing so would slow down execution time noticeably). During a gc (scavenge or global), the stack is scanned and pointers are evaluated. Some do not point to Lisp values. Those are leftovers and can be ignored. But some point to valid Lisp objects. These might be valid live objects but they may also be dead Lisp objects suitable for gc'ing, the pointer on the stack coming from an uninitialized value just added or uncovered.
A conservative approach treats any valid object pointed to from the stack as live. A precise approach determines (by other means) which pointers should be considered and follows only those to objects, which it marks as live. Other pointers are ignored and so do not themselves keep objects alive. Allegro CL uses a hybrid approach: precise where possible and conservative where precision is not possible.
The conservative approach is used when the system considers it likely the function is being debugged (because it has stopped at an unexpected place), or when information on the frame is insufficient to tell whether pointers in it should be considered, or when determining whether the frame should be considered will itself take too much time. The precise approach is used whenever possible and is the typical approach used.
The benefit of the conservative approach is that it is easy to implement, which is why it is what was implemented for many releases. The disadvantages are that objects which are in fact garbage are not gc'ed. A more serious consequence is weak vectors are prevented from being cleared and finalizations are delayed (see Section 10.0 Weak vectors, finalizations, static arrays, etc. for information on weak vectors and finalizations). These consequenecs are mitigated by precise gc'ing of the stack.
Consider the following code snippet:
(let ((base (let ((sarray (make-array 10))) (declare (dynamic-extent sarray)) (if (<= length 10) sarray (make-array length))))) (declare (dynamic-extent base)) <body>)
This code snippet happens to work in 9.0 and earlier, but one of the
declarations is lying to the compiler -- the inner dynamic-extent
declaration (it appears first but is within the outer let form)
-- it is declaring
sarray to have dynamic extent,
meaning that the promise is that the value
sarray will not be used beyond the let form
which created it. But the value is in fact returned from that let
form, thus making a lie out of that dynamic-extent declaration. If
any other dynamic-extent allocations are performed in the body, or if
a tail-call is performed, the array (which is now dead) will be
The reason why this technique works in 9.0 and earlier Lisps is because stack-allocated objects are always allocated in their own separate spaces in the stack, and two stack-allocated objects never overlap. It's as if the dynamic-extent declaration really meant "good for the life of this function". This allows the fudging of dynamic-extent declarations shown above, but it also creates potentially huge wastes of space in the stack, forcing the conservative-gc to scan and validate more slots which increases the risk of dead slots being unnecessarily forwarded.
Now dynamic extent objects on the stack are put in a single, Last In,
First Out (lifo) area and objects can be freed when logically they are
outside the scope of the dynamic extent declarations. That means in
sarray may be overwritten before it is
in fact truly garbage. Care must be excercise to ensure dynamic extent
declarations are proper.
The more compact representation of stack-allocated objects in 10.0 allows even the conservative-gc, when it is used, to scan much fewer slots, and guarantees that only objects that are or were once alive will be scanned (and, of course, the precise-gc only scans what is live).
The process known as reward assessment during a global gc
estimates how much space will be saved by the adjustment and
compactification phases of a global gc. If the estimate is less than
the value of the
parameter, those phases are skipped. This can result in a significant
time saving. Adjustment and compatification moves objects in
oldspace. When an object is moved, all pointers to that object must be
found and updated. Skipping adjustment and compactification means that
pointers to live objects in oldspace are automatically valid and do
not have to be found or updated. Note dead objects in oldspace are
discarded regardless of whether adjustment and compactification is
performed or not.
See the sys:gc-parameter function. When called
:minimum-recovery-bytes as its required argument,
it returns the value of the parameter:
(sys:gc-parameter :minimum-recovery-bytes) RETURNS number of bytes to be recovered if the adjustment and compactification phases of a global gc are to be performed.
That value is setf'able. setf'ing the value to 0:
(setf (sys:gc-parameter :minimum-recovery-bytes) 0)
ensures that those phases are never skipped. That is not recommended in general.
The two pieces of newspace are managed as a stop-and-copy garbage collector system. The two areas are the same size. At any one time, one area is active and the other is not. A newspace area is filled from one end to the other. Imagine, for example, a book shelf. Existing books are packed together on the right side. Each new book is placed just to the left of the leftmost book. It may happen that books already placed are removed, leaving gaps, but these gaps are ignored, with each new book still being placed to the left of the location of the last new book. When the shelf fills up, the other shelf (newspace area) is used. First, all books remaining are moved to the other shelf, packed tight to one side, and new books are placed in the next free location.
So with Lisp objects in newspace areas. All existing objects are packed together on one side of the area and new objects are placed in the free space next to the existing objects, with gaps left by objects which are no longer alive being ignored. When the area fills up, Lisp stops and copies all live objects to the other area, packing them tight. Then the process is repeated.
The process of copying objects from one newspace area to the other is called a scavenge. We will discuss the speed of scavenges below, but scavenges are supposed to be so fast that humans usually barely notice them.
A scavenge happens in one of the following cases:
:allocation :lispstatic-reclaimable, for example, see cl:make-array in implementation.htm), or a weak-vector or finalization will cause aclmalloc space to be freed (see Section 10.0 Weak vectors, finalizations, static arrays, etc.).
The first listed cause is under user-control. The second and third causes are under system control and the resulting scavenge cannot be prevented if the system determines it must occur.
The system keeps track of the age of objects in newspace by counting the number of scavenges that the object has survived. The number of scavenges survived is called the generation of an object. When objects are created, they have generation 1, and the generation is increased by 1 at each scavenge.
Of course, many objects become garbage as time passes. (An object is garbage when there are no pointers to it from any other live object. If there are no pointers to an object, nothing can reference or access it and so it is guaranteed never to be looked at again. Thus, it is garbage.) The theory of a generation scavenging garbage collector is that most objects that will ever become garbage will do so relatively quickly and so will not survive many scavenges.
The problem with a stop-and-copy system is that objects that survive have to be moved and moving objects takes time. If an object is going to be around for a while (or for the entire Lisp session), it should be moved out of newspace to some place where it does not have to be moved (or is moved much less often). This is where the other half of the generation scavenging algorithm comes into play. Once an object has survived enough scavenges, it is assumed to be long-lived and is moved to oldspace. Oldspace is not touched during scavenges and so objects in oldspace are not moved during scavenges, thus saving considerable time over a pure stop-and-copy system.
Part of a scavenge is checking the age (generation) of surviving objects and moving those that are old enough to oldspace. The remaining objects are moved to the other newspace area. The age at which objects are tenured is user-settable. Its initial value is 4 and that seems to work for many applications. We will discuss below how changing that (and many other) settings can affect gc performance.
The process of moving an object to oldspace is called tenuring and the object moved is said to be tenured. At one point, oldspace was also called tenured space and you may see that term occasionally in Allegro CL documents.
Note the assumption: objects that survive a while are likely to survive a long while.
If one could know exactly how long an object is going to survive, one could provide the best possible garbage collection scheme. But that knowledge is not available. Objects are created all the time by different actions and users and even application writers typically do not know what actions create objects or how long those objects will live. Indeed, that information often depends on future events that are hard to control -- such as the behavior of the person running the application.
So the algorithm makes that assumption: if an object survives for a while, it is likely to survive for a long while, perhaps forever (forever means the length of the Lisp session). Of course, for many objects this assumption is wrong: the object may become garbage soon after it is tenured. However, as we said above, scavenges (which are automatic and cannot be prevented by a user although they can be triggered by a user) do not touch oldspace. In order to clear garbage from oldspace, a global garbage collection (global gc) must be done. An interface for automating global gc's is provided in Allegro CL and different interfaces are easy to implement (see below for more information), but the two important points about global gc's are:
The Lisp heap grows upwards (to higher addresses). Oldspaces are at low addresses and newspace occupies addresses higher than any oldspace area. This means that newspace can grow without affecting oldspace and oldspace can grow (usually by creating a new oldspace area) by having newspace move up as far as necessary.
Why might newspace grow? Suppose, for example, a newspace area is 600 Kbytes and you want to allocate a 1 Mbyte array. Newspace has to grow to accommodate this.
Why might oldspace grow? As objects are tenured to oldspace, it slowly fills up. Even with regular global gc's, it can fill up. When it does, newspace moves up and a new old area is created. (New areas are created rather than a single area being expanded for various technical reasons. We discuss below how to reduce sizes dynamically. See Section 3.5 Can other things be changed while running?.)
We will not describe the internal algorithms of the garbage collector because they cannot be changed or modified by users in any way. But let us consider how newspace might be moved, as this might make the process clearer. Suppose the current scavenge is about to move live objects to the high address area. Before anything is moved, Lisp can compute how much space the live objects need, how much space objects waiting to be allocated need, and how much space a new old area needs. From that information, it can compute the highest address of the high address newspace area. It requests from the Operating System that the area be allocated (using malloc or sbrk), and once the Operating System confirms the allocation, starts copying the live objects to that high address, filling toward lower addresses. When all live objects have been moved and new objects are allocated in the high address newspace area, the new oldspace area (if one is required) can be created and the location of the low address newspace area can be determined. Recall that high address newspace area is active so the low address newspace area does not contain anything of importance.
A consequence of what we just said about newspace moving when it has to grow or when a new oldspace area is needed is that the size of the Lisp image can grow while it is running. This is usually normal, indeed what you want. It allows images to start small and grow as much (but no more) than they need. It also allows the same image to run effectively on machines with different configurations.
But, sometimes growth can be unexpected and the image can want to grow to a size larger than the Operating System can handle (usually because there is not enough swap space).
The growth is often necessary, because of the type of application being run. What is important is that the growth be managed and be no more than is really needed.
In earlier releases, space for foreign code loaded into the image, space for foreign objects, and direct calls to malloc all could cause a gap to be placed above newspace. If a new oldspace or a larger newspace was needed, it had to be placed above the gap, causing in some cases a small need for additional space to result in a multimegabyte increase in image size. Now, malloc space is placed away from the new and old spaces and so the Lisp heap (new and old spaces together) are unaffected and can grow incrementally as needed. There is a Lisp heap size specified by the lisp-heap-size argument to build-lisp-image (see building-images.htm). The OS will try to reserve this space when Lisp starts up. If more space is needed, Lisp will request it from the OS but it is possible more space will not be available. If this happens, you might increase the original request.
The space reserved in a running Lisp is reported as 'resrve' on the
'Lisp heap' line of the output of
(room t). If the
heap grows larger than that size, gaps may appear. If you see gaps in
your application, you should consider starting with an image with a
larger heap size.
Application writers and users can control the behavior of the garbage collector in order to make their programs run more efficiently. This is not always easy, since getting optimal behavior depends on knowing how your application behaves and that information may be difficult to determine. Also, there are various paths to improvement, some of which work better than others (but different paths work better for different applications).
One thing to remember is that (unless the image needs to grow larger than available swap space), things will work whether or not they work optimally. You cannot expect optimal gc behavior at the beginning of the development process. Instead, as you gather information about your application and gc behavior, you determine ways to make it work better.
The automated gc system is controlled by switches and parameters (they are listed in Section 5.0 System parameters and switches below). There is not much difference between a switch and a parameter (a switch is usually true or false, a parameter usually has a value) and there probably should not be a distinction, but these things are hard to change after they are implemented. The functions gsgc-switch and gsgc-parameter can be used to poll the current value and (with setf) to set the value of switches and parameters.
The function gsgc-parameters prints out the values of all switches and parameters:
cl-user(14): (sys:gsgc-parameters) :generation-spread 4 :current-generation 4 :tenure-limit 0 :free-bytes-new-other 131072 :free-percent-new 25 :free-bytes-new-pages 131072 :expansion-free-percent-new 35 :expansion-free-percent-old 35 :quantum 32 (switch :auto-step) t (switch :use-remap) t (switch :hook-after-gc) t (switch :clip-new) nil (switch :gc-old-before-expand) nil (switch :next-gc-is-global) nil (switch :print) t (switch :stats) t (switch :verbose) nil (switch :dump-on-error) nil cl-user(15):
gsgc-switch can poll
and set switches while gsgc-parameter can poll and set parameters:
Here we poll and set the
cl-user(15): (setf (sys:gsgc-switch :print) nil) nil cl-user(16): (sys:gsgc-switch :print) nil cl-user(17): (setf (sys:gsgc-switch :print) t) t cl-user(18): (sys:gsgc-switch :print) t cl-user(19):
The gc function can be used to toggle some of the switches.
The system will cause a scavenge whenever it determines that one is necessary. There is no way to stop scavenges from occurring at all or even to stop them from occurring for a specified period of time.
However, you can cause a scavenge by calling the gc function with no arguments:
(excl:gc) ;; triggers a scavenge
You can also cause a scavenge and have all live objects tenured by
calling the gc function with
:tenure, like this
Global gc's (a gc of old and new space) are not triggered automatically (but triggering can be automated). You can trigger a global gc by calling gc with the argument t:
(excl:gc t) ;; triggers a global gc
See section Section 6.0 Global garbage collection for information on other ways to trigger a global gc and ways to automate global gc's.
The function room provides
information on current usage (it identifies oldspaces and newspace and
free and used space in each). Setting the
:verbose switches causes the system to print
information while gc's are occurring. See
Section 5.6 Gc switches and
Section 3.1 How do I find out when scavenges happen?.
(room t) provides the most
information about the current state of memory management. Here is a
(room t) output from a Allegro CL image that has
been doing a fair amount of work. This is from a UNIX machine and was
done immediately after a global gc, so some of the oldspaces have
significant free space.
CL-USER(1): (room t) area area address(bytes) cons other bytes # type 8 bytes each (free:used) (free:used) Top #x106a0000 New #x10598000(1081344) 134:13113 340128:582984 New #x10490000(1081344) ----- ----- 1 Old #x10290000(2097152) 0:0 2095952:0 0*Old #x10000e80(2683264) 0:68273 0:2124792 Tot (Old Areas) 0:68273 2095952:2124792 * = closed old area Root pages: 61 Lisp heap: #x10000000 pos: #x106a0000 resrve: #x10fa0000 Aclmalloc heap: #x64000000 pos: #x64011000 resrve: #x640fa000 Pure space: #x2d1aa000 end: #x2d747ff8 code type items bytes 112: (SIMPLE-ARRAY T) 6586 930920 28.3% 1: CONS 80502 644016 19.6% 8: FUNCTION 8432 520192 15.8% 7: SYMBOL 17272 414528 12.6% 117: (SIMPLE-ARRAY CHARACTER) 2259 259984 7.9% 96: (SHORT-SIMPLE-ARRAY T) 17498 151736 4.6% 18: BIGNUM 2966 139928 4.3% 125: (SIMPLE-ARRAY (UNSIGNED-BYTE 8)) 31 87816 2.7% 12: STANDARD-INSTANCE 3291 52656 1.6% 9: CLOSURE 2301 39688 1.2% 15: STRUCTURE 666 24944 0.8% 127: (SIMPLE-ARRAY (UNSIGNED-BYTE 32)) 9 9920 0.3% 108: (SHORT-SIMPLE-ARRAY CODE) 16 7368 0.2% 10: HASH-TABLE 108 3456 0.1% 17: DOUBLE-FLOAT 120 1920 0.1% 111: (SHORT-SIMPLE-ARRAY FOREIGN) 51 1216 0.0% 16: SINGLE-FLOAT 141 1128 0.0% 118: (SIMPLE-ARRAY BIT) 11 296 0.0% 20: COMPLEX 11 176 0.0% 80: (ARRAY T) 7 168 0.0% 11: READTABLE 8 128 0.0% 123: (SIMPLE-ARRAY (SIGNED-BYTE 32)) 1 88 0.0% 13: SYSVECTOR 3 48 0.0% 85: (ARRAY CHARACTER) 1 24 0.0% total bytes = 3292344 aclmalloc arena: max size free bytes used bytes total 112 3472 112 3584 496 3472 496 3968 1008 2016 2016 4032 2032 0 12192 12192 4080 0 8160 8160 9200 18400 18400 36800 total bytes: 27360 41376 68736 CL-USER(2):
Newspace is divided into two equal size parts (only one of which is
used at any time). There can be numerous oldspaces: two are shown in
the example, but many more are common after Lisp has run for a
while. Oldpsaces are numbered. The gsgc-parameter
:open-old-area-fence takes such a number as an
Section 5.1 Parameters that control generations and tenuring
for information on gsgc-parameters).
The 0th old area in the output is closed, as indicated by the
asterisk. If there are no closed old areas
:open-old-area-fence) returns 0) then no asterisks show
up and the "* = closed old area" note isn't given. When asterisks are
shown, they denote any old areas that are closed. See the discussion
of :open-old-area-fence in
Section 5.1 Parameters that control generations and tenuring
and also the note on closed old areas after the table for information
on closed and open old areas.
Root pages: 61
Root pages contain information about pointers from oldspace to newspace.
Lisp heap: #x10000000 pos: #x106a0000 resrve: #x10fa0000 Aclmalloc heap: #x64000000 pos: #x64011000 resrve: #x640fa000 Pure space: #x2d1aa000 end: #x2d747ff8
The first value is the starting address of the specified heap in memory. The `Pure space' line only appears in Lisps which use a pll file (see pll-file), showing where the pll files is mapped.
In the first two lines, pos (position) is the highest-use location; this is one byte larger than the highest memory used by the lisp. Some operating systems will only commit (assign physical pages) to memory between base (inclusive) and position (exclusive). This is a hexadecimal address value. resrve (reserved) is the number of bytes lisp thinks is reserved to it in virtual memory space. On some operating systems which support it, addresses greater than position, but less than starting location+reserved, will not be overwritten by shared-libraries, other memory mapping operations, etc.
The Lisp heap reserved size is a true limit only for certain free products. With paid license images (and some free products), this value is important only because if the heap grows larger than this limit, gaps in the heap may appear. See Section 1.11 The almost former gap problem for more information. This value is not a limit in any sense on how big the image can grow.
The Aclmalloc heap was called the "C heap" in earlier releases but its name was changed to reflect its real nature. It is the aclmalloc area used for space allocated by the aclmalloc function. That function differs from malloc() in that it ensures that Lisp will remember the location of aclmalloc'ed allocations and preserve it through dumplisp and restarts, thus guaranteeing that aclmalloc addresses remain valid.
More information on aclmalloc and regular malloc():
malloc()space is always started fresh when a Lisp starts up, so addresses used by malloc calls must never be used across dumplisps.
aclmalloc()is aclmalloc. There is no Lisp interface to
malloc()(there is an internal function called excl::malloc, but it calls
aclmalloc(), and not
malloc(), which is why it is not exported). Foreign definitions for
free()can be made in order to gain access to these functions. However, calling
malloc()in a Lisp environment is dangerous, and should only be done after careful consideration of the above.
The type counts are as if printed by print-type-counts:
code type items bytes 112: (SIMPLE-ARRAY T) 6586 930920 28.3% 1: CONS 80502 644016 19.6% 8: FUNCTION 8432 520192 15.8% 7: SYMBOL 17272 414528 12.6% 117: (SIMPLE-ARRAY CHARACTER) 2259 259984 7.9% 96: (SHORT-SIMPLE-ARRAY T) 17498 151736 4.6% 18: BIGNUM 2966 139928 4.3% 125: (SIMPLE-ARRAY (UNSIGNED-BYTE 8)) 31 87816 2.7% 12: STANDARD-INSTANCE 3291 52656 1.6% 9: CLOSURE 2301 39688 1.2% 15: STRUCTURE 666 24944 0.8% 127: (SIMPLE-ARRAY (UNSIGNED-BYTE 32)) 9 9920 0.3% 108: (SHORT-SIMPLE-ARRAY CODE) 16 7368 0.2% 10: HASH-TABLE 108 3456 0.1% 17: DOUBLE-FLOAT 120 1920 0.1% 111: (SHORT-SIMPLE-ARRAY FOREIGN) 51 1216 0.0% 16: SINGLE-FLOAT 141 1128 0.0% 118: (SIMPLE-ARRAY BIT) 11 296 0.0% 20: COMPLEX 11 176 0.0% 80: (ARRAY T) 7 168 0.0% 11: READTABLE 8 128 0.0% 123: (SIMPLE-ARRAY (SIGNED-BYTE 32)) 1 88 0.0% 13: SYSVECTOR 3 48 0.0% 85: (ARRAY CHARACTER) 1 24 0.0% total bytes = 3292344
The aclmalloc arena describes allocation of space for aclmallocs and foreign data. It is divided into chunks of various sizes to allow allocation of requests of various sizes without fragmentation. (Space allocated by aclmalloc is freed by aclfree.)
aclmalloc arena: max size free bytes used bytes total 112 3472 112 3584 496 3472 496 3968 1008 2016 2016 4032 2032 0 12192 12192 4080 0 8160 8160 9200 18400 18400 36800 total bytes: 27360 41376 68736
As a user, or as an application writer, how can you get the garbage collector to work best for you? At first, you do not have to do anything. The system is set up to work as delivered. You will not run out of space, global gc's will happen from time to time (as described below, see Section 6.0 Global garbage collection), the image will grow as necessary, and assuming you do not run out of swap space, everything will work.
Of course, it will not necessarily work as well as it could. As delivered, the garbage collector is set to work best with what we assume is a typical application: objects, none of which are too big, are created as needed. Most objects that survive a while are likely to survive a long while or perhaps forever, and so on. If your application's use of Lisp has different behavior, performance may be suboptimal.
So what to do? One problem is that optimizing gc behavior is a multidimensional problem. Factors that affect it include
Optimization in a multidimensional environment is always complicated.
The first step is always to gather the information necessary to do the tuning. Information like:
Section 3.1 How do I find out when scavenges happen?
Section 3.2 How many bytes are being tenured?
Section 3.3 When there is a global gc, how many bytes are freed up?
Section 3.4 How many old areas are there after your application is loaded?
Section 3.5 Can other things be changed while running?
There are three gsgc switches (these control the behavior of the
garbage collector) that affect printing information about the garbage
(setf (sys:gsgc-switch :print) t)
will cause a short message to be printed whenever a scavenge
happens. Unless the
t, no message will be printed.
switches control the amount of information printed. If the
:stats switch is true, the message contains more
information but the information is compact. If the
:verbose switch is also true, a longer, more easily
understood message is printed.
;; In this example, we cause a scavenge with all flags off, ;; then with :print true, then :print and :stats true, ;; and finally :print, :stats, and :verbose all true. cl-user(5): (gc) cl-user(6): (setf (sys:gsgc-switch :print) t) t cl-user(7): (gc) gc: done cl-user(8): (setf (sys:gsgc-switch :stats) t) t cl-user(9): (gc) gc: E=0% N=866432 T+=3088 A-=0 pfu=0+3 cl-user(10): (setf (sys:gsgc-switch :verbose) t) t cl-user(11): (gc) scavenging...done eff: 0%, new copy: 9472 + tenure: 480 + aclmalloc free: 0 = 9952 Page faults: non-gc = 0 major + 1 minor Page faults: gc = 0 major + 2 minor cl-user(12):
:stats true, the message contains
much more information, but it is coded -- E means Efficiency; N means
bytes copied in newspace; T means bytes copied to oldspace (i.e. bytes
tenured); A- means aclmalloc data freed; pfu means non-gc page faults
and pgu (not shown) means page faults during garbage collection. The
pfu and pfg values may be left out if there were no page faults. With
:verbose also true, the same information is
displayed in expanded form and additional information (about page
faults) is provided.
Efficiency is defined as the ratio of cpu time not associated with gc to total cpu time. Efficiency should typically be 75% or higher, but the efficiencies in the example are low because we triggered gc's without doing anything else of significance.
It is usually desirable to have
:stats true while developing software. This allows you to
monitor gc behavior and see if there seems to be a problem.
That information is shown when the
:stats switches are true, but perhaps the real
question is whether things are being tenured that would be better left
in newspace (because they will soon become garbage). This often
happens when a complex operation (like a compile of a large file) is
being carried out. This question, in combination with the next can
tell you if that is the case.
In the following, copied from above, 0 bytes are tenured in the first gc (T+=0) and 16064 in the second (tenure: 16064):
cl-user(9): (gc) gc: E=17% N=17536 T+=0 A-=0 cl-user(10): (setf (sys:gsgc-switch :verbose) t) t cl-user(11): (gc) scavenging...done eff: 15%, copy new: 1664 + tenure: 16064 + aclmalloc free: 0 = 17728 Page faults: gc = 0 major + 2 minor cl-user(12):
switches are true, the amount of space freed by a global gc is printed
at the end of the report. Here is an example. The form
t) triggers a global gc.
cl-user(13): (gc t) gc: Mark Pass...done(1,583+66), marked 128817 objects, max depth = 17, cut 0 xfers. Weak-vector Pass...done(0+0). Cons-cell swap...done(0+67), 346 cons cells moved Symbol-cell swap...done(17+0), 1 symbol cells moved Oldarea break chain...done(83+0), 40 holes totaling 6816 bytes Page-compaction data...done(0+0). Address adjustment...done(1,400+67). Compacting other objects...done(150+0). Page compaction...done(0+0), 0 pages moved New rootset...done(667+0), 20 rootset entries Building new pagemap...done(83+0). Merging empty oldspaces...done, 0 oldspaces merged. global gc recovered 9672 bytes of old space. gc: E=0% N=1504 T+=0 A-=0 pfg=54+187 cl-user(14):
The next to last line reports on what was recovered from oldspace (9672 bytes). The value is often much higher. It is low in this example because we have not in fact done anything significant other than test gc operations.
There is plenty of other information but we will not describe its meaning in detail. It is typically useful in helping us help you work out complicated gc problems.
The amount of space freed is a rough measure of how many objects
are being tenured that perhaps should be left for a while longer in
newspace. If the number is high, perhaps things are being tenured too
quickly (increasing the value of the
switch will keep objects in newspace longer, as will a larger
The output printed by room
shows the two newspace areas and the various oldspace areas. Here is
an example of room
output. (room takes an
argument to indicate how much information should be displayed). The
following is the output of
(cl:room t), which
causes the most information to be displayed.
cl-user(3): (room t) area area address(bytes) cons other bytes # type 8 bytes each (free:used) (free:used) Top #x569a000 New #x5134000(5660672) 5:95781 597040:4239616 New #x4bce000(5660672) ----- ----- 7 Old #x498e000(2359296) 458:18903 31904:2170416 6 Old #x494e000(262144) 0:1019 0:253648 5 Old #x478e000(1835008) 0:41779 0:1498064 4 Old #x474e000(262144) 0:23437 0:73424 3 Old #x45ce000(1572864) 0:27513 0:1350736 2 Old #x454e000(524288) 0:7133 0:466512 1 Old #x448e000(786432) 0:4076 0:753104 0 Old #x4000d00(4772608) 0:97824 0:3983672 Tot (Old Areas) 458:221684 31904:10549576 Root pages: 158 Lisp heap: #x4000000 pos: #x569a000 resrve: 23699456 Aclmalloc heap: #x54000000 pos: #x54027000 resrve: 1024000 code type items bytes 96: (simple-array t) 76658 3864816 22.8% 108: (simple-array code) 8699 3608136 21.3% 1: cons 314901 2519208 14.9% 99: (simple-array (unsigned-byte 16)) 10938 2242320 13.2% 101: (simple-array character) 38383 1632920 9.6% 8: function 21721 1284216 7.6% 7: symbol 36524 876576 5.2% 107: (simple-array (signed-byte 32)) 264 264336 1.6% 12: standard-instance 14244 227904 1.3% 9: closure 8854 145448 0.9% 98: (simple-array (unsigned-byte 8)) 44 105184 0.6% 97: (simple-array bit) 49 103952 0.6% 15: structure 830 33144 0.2% 100: (simple-array (unsigned-byte 32)) 12 10264 0.1% 10: hash-table 225 7200 0.0% 18: bignum 410 4480 0.0% 16: single-float 505 4040 0.0% 111: (simple-array foreign) 103 2464 0.0% 17: double-float 124 1984 0.0% 64: (array t) 22 528 0.0% 65: (array bit) 13 312 0.0% 13: sysvector 14 224 0.0% 20: complex 12 192 0.0% 11: readtable 7 112 0.0% 69: (array character) 1 24 0.0% total bytes = 16939984 aclmalloc arena: max size free bytes used bytes total 48 3024 48 3072 496 3968 0 3968 1008 4032 0 4032 2032 2032 2032 4064 4080 8160 36720 44880 5104 10208 10208 20416 9200 27600 9200 36800 20464 20464 20464 40928 total bytes: 79488 78672 158160 cl-user(4):
The output shows the two equal size newspace areas, only one of which is being used. It also shows eight oldspaces and provides information about what is in the oldspaces. Then information is printed about other objects such as the number of root pages (a root page keeps information on pointers from oldspace to newspace -- these pointers must be updated after a scavenge), and the locations of the Lisp and C heaps. Then, there is a table showing the types and numbers of objects. Finally, used and available malloc space is displayed.
Yes. The function resize-areas can be used to rearrange things while running. It is typically useful to call this function, for example, after loading a large application. If you know desirable old- and newspace sizes for your application, it is preferable to build an image with those sizes (using the :oldspace and :newspace arguments to build-lisp-image, see building-images.htm for more information). However, you may not know until runtime what the best sizes are, in which case you can call resize-areas on application startup. Be warned that it may take some time.
Another use of resize-areas is when you wish to dump an image (with dumplisp) into which your application has been loaded. You call resize-areas just before dumplisp in that case.
The initial sizes of newspace and oldspace are determined when the image is built with build-lisp-image. See building-images.htm (where build-lisp-image is fully documented -- the page on it is brief to avoid two sources for the same complex discussion) The :newspace argument to build-lisp-image controls the initial size of newspace and the :oldspace argument the initial size of oldspace.
An image dumped with dumplisp inherits new and oldspace sizes from the dumping image. See dumplisp.htm.
resize-areas will restructure old and newspace sizes in a running image.
The garbage collector will automatically resize old and newspace when it needs to. The amount of resizing depends on space required to allocate or move to oldspace live objects, and also on the parameters that relate to sizes.
Gc parameter and switch settings are displayed by the function sys:gc-parameters. All parameters and switches are described in the subsections of this section.
The parameters and switches described under the next set of headings control the action of the garbage collector. You may change them during run time to optimize the performance of the Lisp process. Many parameters and switches values may be set with setf. However, some values should not be changed by you and others cannot be. The descriptions of the parameters say whether you should change their values. By default, the system does automatically increase the generation number. You may find that it is useful to step it yourself at appropriate times with a call to gsgc-step-generation.
There is really no difference between a parameter and a switch
other than the value of switches is typically
nil while parameters
often have numeric values.
The function gc-parameters prints the values of all parameters and switches. gc-switch and gc-parameter retrieve the value of, respectively, a single switch or a single parameter, and with setf can set the value (when appropriate) as follows.
(setf (sys:gc-parameter parameter) value) (setf (sys:gc-switch switch) value)
Switches and parameters are named by keywords. Here are the switches and parameters as printed by sys:gc-parameters:
cg-user(23): (sys:gc-parameters) gc-monitor [version 100] (multithreaded global gc) :helper-threads-requested 4 (limit is 8) :packets-in-use-limit 16,777,215 :minimum-recovery-bytes 8,192 :generation-spread 4 :current-generation 5 :tenure-limit 1 :free-bytes-new-other 131,072 :free-percent-new 25 :free-bytes-new-pages 131,072 :expansion-free-percent-new 35 :expansion-free-percent-old 35 :quantum 32 :open-old-area-fence 0 (switch :auto-step) t (switch :hook-after-gc) t (switch :clip-new) nil (switch :gc-old-before-expand) nil (switch :next-gc-is-global) nil (switch :print) t (switch :stats) t (switch :verbose) nil (switch :dump-on-error) nil cg-user(24):
The first three parameters relate to the generation number and when
objects are tenured. Please note that of the three parameters, you
should only set the value of
The fourth parameter, which is setf'able, allows closing off some old
areas, meaning that no objects will be tenured to them. Old areas are
now numbered, allowing for some to be closed off.
| ||The value of this parameter is a 16 bit unsigned integer. New objects are created with this generation tag. Its initial value is 1, and it is incremented when the generation is stepped. The system may change this value after a scavenge. Users should not set this value. Note: Both the current generation number and the generation of an individual object are managed in a non-intuitive way. While it is conceptually correct that the generation number increases, the actual implementation works quite differently, often resetting the generation number toward 0.|
| ||The value of this parameter is a 16 bit integer. During a scavenge, objects whose generation exceeds this value are not tenured and all the rest are tenured. Users should not set this value. Its initial value is 0, and it is constantly being reset appropriately by the system.|
| ||The value of this parameter is the number of distinct generations that will remain in newspace after garbage collection. Note: objects that are marked for tenuring and objects that are to stay in newspace permanently do not belong to a specific generation. Setting the value of this parameter to 0 will cause all data to be tenured immediately. This is one of the most important parameters for users to set. Its initial value is 4 and its maximum effective value is 25.|
| ||The value of this parameter is always a
non-negative integer which is the number of the oldest old area that
is open (not closed). Old areas are numbered with 0 as the oldest
This parameter is setfable, either with the number of
the old-area that is desired to be the first open-old-area, or with
a negative number, for which the old-areas are counted backward to
set the fence. For example,
See the note on closed old areas just after this table for more information.
Old areas can be marked as closed. No objects are newly tenured into a closed old-area; it is as if the area is full. Also, no dead object in a closed old area is collected while the area is closed, and data pointed to by that object is also not collected.
See the description of the
the table just above for details on how to specify old areas as
The intended usage model for closing old areas is this: a programmer with an application, such as a VAR, will load up their application, perform a global-gc and possibly a resize-areas, and then close most of the old-areas, leaving room for their users' data models to be loaded into the open-areas. When the user is done with the data model, it can be thrown away and a fast global-gc performed, making way for the next data model.
The following parameters control the minimum size of newspace and when
the system will allocate a new newspace. At the end of a scavenge, at
:free-bytes-new-other bytes must be free, and at
:free-percent-new percent of newspace must be
free. If these conditions are not met, the system will allocate a new
newspace, large enough for these conditions to be true after
allocating the object that caused the scavenge, if there is
one. (Unless explicitly called by the user, a scavenge occurs when the
system is unable to allocate a new object.) Note that there is no
system reason why there are two parameters,
:free-bytes-new-other -- differences were
anticipated in the original specification but none was never
implemented. The two parameter values are added to get total free
||The value of this parameter is a 32-bit integer which represents the minimum amount of space (in 8 Kbyte pages) that will be requested for a new new or old space and the granularity of space requested (that is space will be requested in multiples of :quantum pages). Its initial value is 32. This parameter value is overshadowed by the other size-related parameters described immediately below, and for that reason, we do not recommend that you change this value.|
||This is one of the parameters which determine the minimum free space which must be available after a scavenge. Its initial value is 131072.|
||This is one of the parameters which determine the minimum free space which must be available after a scavenge. Its initial value is 131072.|
||This parameter specifies the minimum fraction of newspace which must be available after a scavenge, or else new newspace will be allocated. Its initial value is 25.|
The next two parameters control how large a new newspace (and a new oldspace) will be. If newspace is expanded or a new oldspace is allocated, then at least the percentage specified by the appropriate parameter shall be free, after, in the case of newspace, the object that caused the scavenge has been allocated, and after, in the case of oldspace, all objects due for tenuring have been allocated. There are different concerns for the newspace parameter and the oldspace parameter.
Let us consider the oldspace parameter first. In the case where no
foreign code is loaded, then oldspaces are carved out of newspace, and
newspace grows up into memory as needed. If each new oldspace is just
large enough, the next time an object is tenured, another oldspace,
again just large enough, will be created, and the result will be a
bunch of small oldspaces, rather than a few larger ones. This problem
will not occur if there is foreign code, since some oldspaces will be
as large as previous newspaces. If the function room shows a bunch of
little oldspaces, you might try increasing the
:expansion-free-percent-old parameter to cure the
problem. However, resize-areas can be used instead to coalesce
the oldspaces into one.
The newspace parameter is more complicated, since newspace can grow
incrementally (assuming growth is not blocked by foreign code). Since
growing newspace takes time, you want to ensure that when newspace
grows, it grows enough. Therefore, it is essential that
:expansion-free-percent-new be larger than
:free-percent-new. Otherwise, you might find
newspace growing just enough to satisfy
:free-percent-new, and then having to grow again at
the next scavenge, since allocating a new object again reduced the
free space below the
||At least this percentage of newspace must be free after allocating new newspace. The system will allocate sufficient extra space to guarantee that this condition is met. Its initial value is 35.|
||At least this percentage of space must be free in newly allocated oldspace (note: not the total oldspace). Its initial value is 35.|
A global garbage collection can take an appreciable amount of
time. The grbage collector can estimate how many bytes will be
recovered in oldspace. If that number is less than the value of the
:minimum-recovery-bytes parameter, then some phases
of the global gc (the adjustment and compaction phases) will be
skipped. This will leave garbage in oldspace that would otherwise be
cleared but can also save significant time. The initial value of the
:minimum-recovery-bytes parameter parameter is
8192. setf can be used
with sys:gc-parameter to
change the value. Setting it to 0 means every global gc will be full.
cg-user(27): (sys:gc-parameter :minimum-recovery-bytes) 8192 cg-user(28): (setf (sys:gc-parameter :minimum-recovery-bytes) 0) 0 ;; All global gc's will be full cg-user(29): (sys:gc-parameter :minimum-recovery-bytes) 0 cg-user(30): (setf (sys:gc-parameter :minimum-recovery-bytes) 16384) 16384 ;; global gc's will be less likely to be full ;; compared to default value cg-user(31): (sys:gc-parameter :minimum-recovery-bytes) 16384 cg-user(32):
:helper-threads-requested parameter tells
the system how many threads it should use, if possible, for garbage
collection. This is in addition to the basic gc thread always
used. Broadly speaking the more the faster, but in SMP Lisps, you may
not want more threads than processors, and, if some processors are
dedicated to other tasks (like running the video), you may want fewer
requested than processors on your machine. setf can be used with sys:gc-parameter to set this parameter:
cg-user(35): (sys:gc-parameter :helper-threads-requested) 4 cg-user(36): (setf (sys:gc-parameter :helper-threads-requested) 6) 6 cg-user(37): (sys:gc-parameter :helper-threads-requested) 6 cg-user(38):
:helper-thread-limit parameter is not
setf'able. It is the
maximum number of helper threads that can be used. Its value is
determined by internal characteristics of the garbage collector and is
unrelated to hardware. Therefore the limit may be larger than the
number of cores even though there is no benefit to requesting more
helpers than the number of cores minus 1.
:helper-threads-requested must be less than or
:helper-thread-limit. This value is
reported in the
sys:gc-parameters output as
part of the
(":helper-threads-requested 4 (limit is 8)"). Pass
:helper-thread-limit to sys:gc-parameter to get the value
:helper-thread-count is the number of
threads actually being used. This is just a report and the value is not
:packets-in-use-limit parameter is a
system value which should not be set except as part of debugging a
:packets-required value is
not setf'able and is
a reported value which may be useful for debugging purposes but can
otherwise be ignored.
There are several switches which control the action of gsgc. The
value of a switch must be
nil. The function gc-switch takes a switch name as an
argument and returns its value or with setf, sets its
also prints out their values. The switches can be set by evaluating
(setf (sys:gc-switch switch-name) nil-or-non-nil)
||If this switch is set true, then before
expanding oldspace, the system will do a global garbage collection (that is, it will
gc oldspace) to see if the imminent expansion is necessary. If enough space is free after
the garbage collection of oldspace the expansion will not occur. Initially
|If true, print a message when a gc occurs. Can be set by
length of the message is determined by the next two switches. If
||If true and
||If true, make the message printed (when :print is true) more
||This is the most important of the switches. If true, which is its initial value, gsgc-step-generation is effectively called after every scavenge. Thus (with the default :generation-spread) an object is tenured after surviving four scavenges.|
||If this switch is true, the function object
bound to the variable
If this switch is set true, the next gc will
be a global gc (that is both newspace and oldspace will be gc'ed). After the global gc,
the system will reset this switch to
The difference between setting this switch and causing a global gc explicitly with the function excl:gc is that setting this switch causes the system to wait until a scavenge is necessary before doing the global gc while calling the function causes the global gc to occur at once. The system uses this switch under certain circumstances.
The scavenger maintains a new logical pool of memory in
newspace called `reserved'. When the
If :print and :verbose are both true, information about the action triggered by this switch is printed. The information refers to `hiding' (moving space to the reserved bucket) and `revealing' (moving space to the free bucket).
If this switch is set true and the operating system on which Allegro CL is running supports it, then physical memory pages that are no longer used after a garbage collection are given back to the operating system in such a fashion that paging is improved.
Specifically, when this switch is true and the currently used half of newspace is copied to the unused half, the following things are done with the previously-used memory area: (1) the operating system is advised to ignore the page reference behavior of those addresses, and (2) the memory is unmapped and then is remapped, after being filled with zeros. The zero filling is necessary for security reasons, since the memory given back to the operating system will potentially be given to another process that requests virtual memory, without first being cleared. If it were not for (2), then remapping would always be advantageous and there would be no switch to control this behavior. As it is, there may be certain situations where zero filling will be too expensive, especially on machines which have a very large amount of physical memory and the decrease in locality does not effect the runtime performance of the Allegro CL application, or where the mmap() implementation is flawed.
If this switch is set true, then a core dump is automatically taken when an internal garbage collection error occurs. The core dump will fail, however, if (1) there is a system imposed limit on the size of a core dump and dumping the image would exceed this limit or (2) there is some other system impediment to dumping core, such as the existence of a directory named core. We assume that you can prevent the second limitation. Here are a few more words on the first limitation. In the C shell, the limit command and its associates can be used to set a higher limit or no limit for the maximum size of a core dump.
If the value of this switch is
||This switch should only be set if instructed as part of a problem report.|
These examples show the effect on gc messages of
nil, a message like the following is printed during a
:verbose is also true but
the message is:
:stats is true and
nil, the message is similar to:
gc: E=34% N=30064 T+=872 A-=0 pfu=0+101 pfg=1+0
The same message with
:verbose true would be:
scavenging...done eff: 9%, new copy: 148056 + tenure: 320 + aclmalloc free: 0 = 148376 Page faults: non-gc = 0 major + 1 minor
abbreviations are used. Their meanings are explained when
:verbose is true. T or Tenure
means bytes tenured to oldspace. A and alcmalloc free
refer to malloc space (see aclmalloc).
E or eff. is efficiency: the ratio of non-gc time and all time (the efficiency is low in our example because we forced gc's in order to produce the example; as we say elsewhere, efficiencies of less than 75% are a cause for concern when the gc is triggered by the system).
The copy figures are the number of bytes copied within newspace and to oldspace.
X means "expanding", so XO means "expanding oldspace" and XN means "expanding newspace". XMN means "expanding and moving newspace".
Page faults are divided between user (pfu or non-gc) caused and gc (pfg or gc) caused. See the Unix man page for getrusage for a description of the difference between major and minor page faults.
Here are a couple of more examples (with
on and off in a fresh Lisp each time):
cl-user(1): (gc :print) t cl-user(2): (setf (sys:gc-switch :verbose ) t) t [...] cl-user(7): (defconstant my-array (make-array 10000000)) scavenging...expanding new space...expanding and moving new space...done eff: 36%, copy new: 7533984 + old: 85232 = 7619216 Page faults: non-gc = 1 major + 0 minor my-array ;; And in a fresh image with :verbose off: cl-user(1): (gc :print) t cl-user(2): (defconstant my-array (make-array 10000000)) gc: XN-XMN-E=32% N=7522488 T+=85632 A-=0 pfu=4+0 my-array
|Function or variable||Arguments of functions||Brief Description|
|gsgc-step-generation||Calling this function, which returns the new value of
:current-generation, increases the current generation number and, if necessary, the value
|gc||&optional action||Called with no arguments, perform a scavenge; called with
|print-type-counts||&optional (location t)||Prints a list of quantities and sizes of lisp objects in the specified location in the heap, along with type names and type codes of each object type printed. See the print-type-counts page for location details.|
|lispval-storage-type||object||Returns a keyword denoting where object is stored. See the lispval-storage-type page for interpretation of the returned value and examples. (In earlier releases, the function pointer-storage-type performed this function. It is still supported, but its use is deprecated. lispval-storage-type is more flexible and should be used instead.)|
|resize-areas||&key verbose old old-symbols new global-gc tenure expand sift-old-spaces pack-heap||This function resizes old and newspaces, perhaps coalescing oldspaces, according to the arguments. See the resize-areas page for details.|
||If the gsgc switch :hook-after-gc is true,
then the value of this symbol, if true, will be funcalled immediately
after a scavenge. See the description of
|gc-after-c-hooks||Returns a list of addresses of C functions that will be called after a gc. See gc-after-c-hooks for details.|
|gc-before-c-hooks||Returns a list of addresses of C functions that will be called before a gc. See gc-before-c-hooks for details.|
In a global garbage collection (global gc), objects in oldspace are garbage collected. Doing so frees up space in oldspace for newly tenured objects. Global gc's are time consuming (they take much longer than scavenges) and they are not necessary for Lisp to run.
The effect of never doing a global gc is the Lisp process will slowly grow larger. The rate of growth depends on what you are doing. The costs of growth are that the paging overhead increases and, if the process grows too much, swap space is exhausted, perhaps causing Lisp to stop or fail.
You have complete control over global gc's. The system will keep track of how many bytes have been tenured since the last global gc. You can choose one of these options for global gc:
The function that records how many bytes have been tenured since
the last global gc is the default value of the variable
*gc-after-hook*. If you set
that variable (whose value must be a function or
nil or a function
that does not keep records of bytes tenured, you will not get the
behavior described here. (See the description of
information on defining a function that does what you want and records
bytes tenured correctly.)
has as its value its initial value or a function that records bytes
tenured correctly, global gc behavior is controlled by the global
*tenured-bytes-limit* is used in conjunction
*global-gc-behavior*. The number of bytes
tenured (moved to oldspace) since the last global gc is remembered and
*global-gc-behavior* depends on when
The tenuring macro causes the immediate tenuring (moving to oldspace) of all objects allocated while within the scope of its body. This is normally used when loading files, or performing some other operation where the objects created by forms will not become garbage in the short term. This macro is very useful for preventing newspace expansion.
It is useful if possible to provide some sort of cue while garbage collections are occurring. This allows users to know that a pause is caused by a gc (and not by an infinite loop or some other problem). Typical cues include changing the mouse cursor, printing a message, or displaying something in a buffer as Emacs does when the emacs-lisp interface is running.
Unfortunately, providing such a cue for every scavenge is a difficult problem and if it is done wrong, the consequences to Lisp can be fatal. However, we have provided an interface for the brave. The functions gc-before-c-hooks and gc-after-c-hooks return settable lists of addresses of C functions to be called before and after a gc.
Luckily, scavenges are usually fast and so failing to provide a cue may not be noticeable. Global gc's, however can be slow but it is possible to provide a cue for a global gc even without using C functions. There are two strategies: (1) determine when a global gc is necessary and either schedule it when convenient or warn the user that one is imminent; (2) do not give advance warning but provide a cue when it happens. Looking at these examples, you can probably craft your own method of warning or cue.
Note that in these examples, we replace the value of
*gc-after-hook* with a new
value, destroying the current value (which provides for the default
automated behavior). The default function is named by the (internal)
One way to implement the first strategy is to set a flag when a global gc is needed and then have code that acts on that flag. This code can be run at your choosing -- but be sure that it is run at some point. You might do this:
(defvar *prevent-gc-recursion-problem* nil) (defvar *my-gc-count* 0) (defun my-gc-after-hook (global-p to-new to-old eff to-be-alloc) (declare (ignore eff to-new to-be-alloc)) (when (atomic-conditional-setf *prevent-gc-recursion-problem* t nil) (if global-p (setq *my-gc-count* 0) (progn (setq *my-gc-count* (+ *my-gc-count* to-old)) (when (> *my-gc-count* excl:*tenured-bytes-limit*) ;; (<change the cursor, print a warning, whatever>) (gc t) ;; (<reset the cursor if necessary, do whatever else>) (setq *my-gc-count* 0)))) (setq *prevent-gc-recursion-problem* nil)))
nilwhen the function executes.
;; (<change the cursor, print a warning, whatever>) ;; (<reset the cursor if necessary, do whatever else>)
with whatever code you want, but be careful that there is no possibility of waiting (for user input, e.g.) or going into an infinite loop.
The following list contains information and advice concerning gsgc. Some of the information has already been provided above, but is given again here for emphasis.
nilunless some other method of stepping the generation is enabled (including specific action by you). If objects are not tenured, newspace will grow, filling up with long-lived objects, and performance will degrade significantly.
:tenure(which will cause all live objects to be tenured) or with the tenuring macro. There is no way to prevent a specific object from ever being tenured except by disabling generation stepping and thus preventing all objects from being tenured.
It is not easy to cause a gsgc error. Such errors are usually catastrophic (often Lisp dies either without warning or with a brief message that some unrecognizable object was discovered). Once the garbage collector becomes confused, it cannot be straightened out.
Such errors can be caused when Lisp code is compiled with the compiler
optimizations set so that normal argument and type checking is
disabled. For example, if a function is compiled with the values of
speed and safety such that the compiler:verify-argument-count-switch
nil, and that function is passed the wrong
number of arguments (usually too few), it can trigger a fatal gsgc
error. Before you report a gsgc error as a bug (and please do report
them), please recompile any code where checking was disabled with
settings of speed and safety which allow checking. See if the error
repeats itself. See Declarations and optimizations in
compiling.htm for more information on compiler
optimization switches and values of speed and safety.
Garbage collector errors may also be caused by foreign code signal handlers. Note that foreign code signal handlers should not call lisp_call_address or lisp_value. See foreign-functions.htm for more information on signals.
See the information on the
Section 5.6 Gc switches.
The Allegro CL image will grow as necessary while running. If it needs to grow and it cannot, a storage-condition type error is signaled (storage-condition is a standard Common Lisp condition type). While these errors might arise from insufficient swap space, the typical cause is a conflict in the virtual address space. That is, something else (a program or a library location) has grabbed virtual address space in a range that Lisp needs to grow the heap. (Allegro CL does not allow discontinuous virtual address ranges.)
Whatever the cause, the error is triggered by a request for space which cannot be fulfilled. Here we show the error when a largish array is created (this example is contrived in order to show the error: a request for such an array does not typically cause a problem).
CL-USER(25): (setq my-array (make-array 100000)) Error: An allocation request for 483032 bytes caused a need for 12320768 more bytes of heap. The operating system will not make the space available. [condition type: STORAGE-CONDITION]  CL-USER(26):
A global gc may free up enough space within Lisp to continue without growing. Killing processes other than Lisp may free enough space for Lisp to grow. But it may be the case that other allocations of virtual address space conflicts with Lisp usage. Please contact Franz customer support for assistance in determining whether this is the case if the problem persists.
You trigger a global gc by evaluating (see gc):
When the garbage collector gets confused, usually by following what it believes to be a valid pointer, but one that does not point to an actual Lisp object, Lisp fails with a two part message. The first part is a brief description of the specific problem. The second part is a general statement about the gc failures and the fact they cannot be recovered from, along with an opportunity of get a core file (which may be useful for later analysis). Here are some examples of the second part:
The gc's internal control tables may have been corrupted and Lisp execution cannot continue. [or] The internal data structures in the running Lisp image have been corrupted and execution cannot continue. [then] Check all foreign functions and any Lisp code that was compiled with high speed and/or low safety, as these are two common sources of this failure. If you cannot find anything incorrect in your code you should contact technical support for Allegro Common Lisp, and we will try to help determine whether this is a coding error or an internal bug. Would you like to dump core for debugging before exiting(y or n)?
Here are some examples of the first part:
system error (gsgc): Unknown object type at (0xc50ec9a) system error (gsgc): Object already pointing to target newspace half: 0x42c43400 system error (gsgc): Scavenger invoked itself recursively.
As the text says in the second part, there is no recovery.
The causes of such errors can be one or more of the following:
Diagnosing and fixing the problem can be difficult. Here are some initial steps to take where possible:
(gc :print), see gc.)
If you cannot quickly determine the cause of the problem and a solution for it, contact Franz Inc. technical support at firstname.lastname@example.org. Be sure to include the output of a call to print-system-state, and provide any information about foreign code, optimizations, etc. that you can. We may ask you for a core file (which, as said above, can be optionally generated when there is a failure).
A Lisp object becomes garbage when nothing points to or references it. The way the garbage collector works is it finds and identifies live objects (and, typically, moves them somewhere). Whatever is left is garbage. Finalizations and weak vectors allow pointers to objects which will not, however, keep them alive. If one of these pointers exists, the garbage collector will see the item and (depending on the circumstances), either keep it alive (by moving it) or abandon it.
It is useful to distinguish two actions of the garbage collector. When the only pointers to an object are weak pointers or finalizations, the garbage collector follows those pointers and `identifies the object as garbage'. If it decides not to keep the object alive, it `scavenges the object away'. Note that any Lisp object can be scavenged away - that just means that the memory it used is freed and (eventually) overwritten with new objects. Only objects pointed to by a weak vector or a finalization can be identified as garbage, however. Live objects are not garbage and objects with nothing pointing to them are not even seen by the garbage collector.
The function lispval-storage-type applied to an object returns a keyword providing information about the storage type of the object. Weak vectors and finalizations are identified by this function which can be used to test whether an object is a weak vector.
Weak arrays are created with the standard Common Lisp function make-array (using the non-standard weak keyword argument) or (if a vector is desired) with the function weak-vector. When you create a weak array by specifying a true value for the weak keyword argument to make-array, you cannot also:
:displaced-tokeyword argument (weak arrays cannot be displaced into other arrays).
:new(the non-standard allocation keyword argument allows specifying where the arrays will be created, with choices including foreign space, non-gc'ed Lisp space, or the Lisp heap, called for by
:new, which is also the default).
When make-array successfully accepts a true value for the weak keyword argument, the object that is weak is always the underlying simple-vector; if the resultant array is non-simple or is multidimensional, then the array itself is not marked as weak, but objects in the array will still be dropped by the gc when not otherwise referenced, because the simple-array that is the data portion of the array is itself weak.
See the discussion of extensions to make-array (in implementation.htm) for further details on the Allegro CL implementation of make-array and the weak keyword argument.
As we said in the brief description above, the most important feature
about weak arrays is that being pointed to by a weak array does not
prevent an object from being identified as garbage and scavenged
away. When an object is scavenged away, the entry in a weak array that
points to the object will be replaced with
Weak arrays allow you to keep track of objects and to discover when
they have become garbage and disposed of. An application of weak
arrays might be determining when some resource external to Lisp can be
flushed. Suppose that all references to a file external to Lisp are
made through a specific pathname. It may be that once all live
references to that pathname are gone (i.e. the pathname has become
garbage) the file itself is no longer needed and it can be removed
from the filesystem. If you have a weak array which points to the
pathname, when the reference is replaced by
nil by the garbage collector, you can tell the system
to kill the file.
It is important to remember that objects which have been tenured will not be identified as garbage unless a global gc is performed. If you use weak arrays, you should either arrange that global gc's are done regularly or that you do an explicit global gc before checking the status of an element of the weak array.
We provide a simple example of weak vectors (weak arrays differ from weak vectors only in having more dimensions) and finalizations in Section 10.3 Example of weak vectors and finalizations.
Weak hashtables are also supported. See implementation.htm for information on an extension to make-hash-table that creates weak hashtables.
The function lispval-storage-type applied to an object returns a keyword providing information about the storage type of the object. Weak vectors and finalizations are identified by this function which can be used to test whether an object is a weak vector.
A finalization associates an object with a function and optionally
queue. If the object
is identified as garbage by the garbage collector, either the function
is called with the object as its single argument before the object is
scavenged away (if there is no associated queue) or a list consisting
of the function and the object is placed on the queue. In the latter
case, no further action is taken by the system. The program must apply
the function call-finalizer at its convenience.
Multiple finalizations can be scheduled for the same object; all are run or placed on queues if and when the gc identifies the object as garbage. The order of the running of multiple finalizations is unspecified.
The timing is important here. While the garbage collector is running, nothing else can be done in Lisp. In particular, functions (except those internal to the garbage collector itself) cannot be called. Therefore, the process of running a finalization and the eventual reclamation of the finalized object occurs over two invocations of the garbage collector. During the first gc, the object is identified as garbage. The garbage collector sees the finalization and so, instead of immediately scavenging the object away, it keeps it alive and makes a note to call the finalization function or enqueue the list of the function and the object as soon as it finishes the current scavenge (or global gc). The finalization is removed from the object after the function is run so that during a subsequent garbage collection, the object is either garbage or, if a weak vector points to it, again identified as garbage and (since it no longer has a finalization) scavenged away. See the example in Section 10.3 Example of weak vectors and finalizations.
A finalization is not a type of Lisp object. Finalizations are implemented as vectors (but that may change in a later release). You should not write any code which depends on the fact that finalizations are implemented as vectors.
You schedule a finalization with the function schedule-finalization. The function unschedule-finalization removes the finalization from the object. Finalizations are only run once, either immediately after the garbage collection which identified the object as garbage or by the program. The object is not scavenged away during the garbage collection in which it is identified as garbage (since it must be around to be the argument to the finalization function).
A reference from a weak vector or hash-table is not a reference from the point of view of the garbage collector. Consequently, a weak vector or hash-table can contain a pointer to an object on which a finalization is being called, or has already been called. If it is important for an application to know when an object has been finalized, the application code must mark the object in some way.
Without extra care on the part of the programmer, finalization functions can run on any thread. In an SMP environment, this can introduce contention for objects found via weak vectors or hash-tables. Programmers should allow for this, using atomic update operations as appropriate. See smp.htm for information on atomic operations.
On Windows, load the file after the transcript example. Typing the example into the Debug window in the Integrated Development Environment on Windows does not work as described, because tools in the IDE cause pointers to the objects to be preserved, and, as a result, the example does not work as described. On Unix, you can type the example directly to a prompt.
;; This example can be typed to a prompt on a UNIX platform. ;; We define a weak vector and a function to be called by finalizations. CL-USER(10): (setq wv (weak-vector 1)) #(nil) CL-USER(11): (defun foo (x) (format t "I am garbage!~%")) foo ;; We create an object and put it in the weak vector: CL-USER(12): (setq a (cons 1 2)) (1 . 2) CL-USER(13): (setf (aref wv 0) a) (1 . 2) ;; We schedule a finalization. CL-USER(14): (schedule-finalization a 'foo) #((1 . 2) foo nil) ;; We remove the reference to the cons by setting the value of A to NIL. CL-USER(15): (setq a nil) nil ;; We evaluate unrelated things to ensure that the object ;; is not the value of *, **, or ***. CL-USER(16): 1 1 CL-USER(17): 2 2 CL-USER(18): 3 3 ;; We trigger a scavenge. The finalization function is called. ;; Note: an automatic gc might occur while you are entering the ;; forms shown. If it does, you might see the message printed by ;; our finalization sooner. CL-USER(19): (gc) I am garbage! ;; At this point, the weak vector still points to the object: CL-USER(20): wv #((1 . 2)) ;; The next gc causes the value in the weak vector to ;; be changed. CL-USER(21): (gc) CL-USER(22): wv #(nil)
On Windows, put this in a file finalization.cl:
(in-package :cg-user) (setq wv (weak-vector 1)) (defun foo (x) (format t "I am garbage!~%")) (setq a (cons 1 2)) (setf (aref wv 0) a) (schedule-finalization a 'foo) (setq a nil)
and do this in the Debug Window (the transcript shows the I am
garbage! message occurring after
evaluated at cl-user(11) but if a gc was triggered by the
loading of finalization.cl, you may see instead
the I am garbage! message sooner):
cl-user(10): :ld finalization.cl Loading finalization.cl cl-user(11): wv #((1 . 2)) cl-user(12): (gc) I am garbage! cl-user(13): 1 1 cl-user(14): 2 2 cl-user(15): 3 3 cl-user(16): (gc) cl-user(17): wv #(nil)
The data in a static array is placed in an area that is not garbage collected. This means that the data is never moved and, therefore, pointers to it can be safely stored in foreign code. (Generally, Lisp objects are moved about by the garbage collector so a pointer passed to foreign code is only guaranteed to be valid during the call. See foreign-functions.htm for more information on this point.) Only certain element types are permitted (listed below). General arrays (whose elements can be any Lisp object) cannot be created as static arrays.
The primary purpose of static arrays is for use with foreign code. They may also be useful with pure Lisp code but only if the array is needed for the whole Lisp run and the array never has to be significantly changed. The problem is the array is not garbage collected and there is no way (within Lisp) to free up the space.
Static arrays are returned by make-array with the (Allegro CL-specific) allocation keyword argument specified. The data is stored in an area which is not garbage collected but the header (if any) is stored in standard Lisp space. The following element types are supported in static arrays.
bit (signed-byte 8) (unsigned-byte 4) (unsigned-byte 8) (signed-byte 16) (unsigned-byte 16) (signed-byte 32) (unsigned-byte 32) character single-float double-float (complex single-float) (complex double-float)
See implementation.htm for information on this extension to make-array. See lispval-other-to-address for information on freeing static arrays.
:gcpath module contains two operators: the
function gcpath:find-gcpath, which finds a path from a Lisp
root pointer to an object; and the
macro gcpath:collected-newstuff, which identifies new objects
of particular types created when evaluating a form. Here are the
Arguments: &optional vector &key slot result weak (interpret t)
This looks for a path from one of Lisp's root pointers to a particular object. This allows users who belive an object is garbage but note it is not garbage collected to figure out why the object is still live.
The target object can be specified in any of several ways.
(gcpath:find-gcpath v :slot 5)in which case the target object would be
(aref v 5).
(gcpath:find-gcpath v)in which case the target is the first non-
(gcpath:find-gcpath ht)in which case the target is the first non-
nil, non-empty, non-deleted key in
ht, which must be a hash-table.
fis a finalization. In this case the target is the object
fis a finalization for.
(gcpath:find-gcpath). This is used after a previous invocation of find-gcpath has found a path too long to fit in the work buffer. In this case an object part way along the path is chosen as a temporary root, a path from that temporary root is displayed, and the temporary root saved to be used as the default target in a subsequent find-gcpath.
The result argument designates a vector in which the path is built. It can be omitted to get a newly-allocated vector of default size. It can be an integer, specifying the length for the newly-allocated vector. It can be a simple-vector of at least the default size which can be panified (and it will be, as part of the operation).
If the weak argument is
nil, then the vector
argument can be a non-weak vector, but will be treated as a weak
vector during the path-tracing.
If the interpret argument is omitted or
nil, then the function returns no values,
but formats an interpreted display of the path. If the interpret
argument is specified as
nil, then no
interpretation is printed, and the function returns the result vector.
;; We create an object, schedule a finalization on it, and ;; apply FIND-GCPATH to the finalization. ;; We first run it in Allegro CL running uder emacs on a Linux machine: cl-user(78): (setq a (make-array 10 :initial-element 0)) #(0 0 0 0 0 0 0 0 0 0) cl-user(79): (schedule-finalization a #'(lambda (x) (format t "I am garbage says ~S!~%" x))) #(#(0 0 0 0 0 0 0 0 0 0) #<Interpreted Function (unnamed) @ #x10001234b42> nil) cl-user(80): (setq f *) #(#(0 0 0 0 0 0 0 0 0 0) #<Interpreted Function (unnamed) @ #x10001234b42> nil) cl-user(81): (gcpath:find-gcpath f) Target is the global value of a which is an interned symbol ;; This is pretty much what is expected. ;; When we set the value of A to NIL, and try again, ;; we get: cl-user(86): (gcpath:find-gcpath f) I am garbage says #(0 0 0 0 0 0 0 0 0 0)! No tether found cl-user(87): ;; Running the same example on a Windows machine using the IDE ;; we note a completely different gcpath is found, involving various ;; IDE tools which remember evaluation history etc. It is things ;; like that which may be keeping alive objects which are otherwise ;; presumed to be dead. ;; cg-user(17): (setq a (make-array 10 :initial-element 0)) #(0 0 0 0 0 0 0 0 0 0) cg-user(18): (schedule-finalization a #'(lambda (x) (format t "I am garbage says ~S!~%" x))) #(#(0 0 0 0 0 0 0 0 0 0) #<Interpreted Function (unnamed) @ #x2027e5f72> nil) cg-user(19): (setq f *) #(#(0 0 0 0 0 0 0 0 0 0) #<Interpreted Function (unnamed) @ #x2027e5f72> nil) cg-user(21): (find-gcpath f) Target is element 2 in a list which is element 4 of a simple-vector(20) which is the instance data of #<outline-item (:form (setq a (make-array 10 :initial-element 0)) #(0 0 0 0 0 0 0 0 0 0)) closed @ #x2028ff932> which is element 3 in a list which is element 6 of a simple-vector(92) which is the instance data of #<history-dropping-outline :history-outline @ #x201b61512> which is element 0 in a list which is element 76 of a simple-vector(108) which is the instance data of #<listener-toolbar :listener-toolbar in IDE GUI @ #x201b6a3e2> which is element 37 of a simple-vector(112) which is the instance data of #<multi-picture-button-pane :listener-toolbar-multipic in IDE GUI @ #x2026c7092> which is element 627 of a simple-vector(787) which is the values vector of a hash-table which is element 4 of a simple-vector(14) which is the instance data of #<cg.base::cg-system @ #x201a569e2> which is the global value of *system* which is an interned symbol cg-user(22):
Arguments: ( &optional (type #()) (size 1000)) &body body
The macro evaluates body, using a weak-key hash-table to record all the objects of the designated type before evaluating the body code. After the body has been evaluated, all the objects of that type are again scanned, and those which are not in the hash-table are collected into a newly-allocated weak simple-vector which is returned as the value of the macro form.
The type to be tracked defaults to
simple-vector. Another type can be designated by providing the
optional type argument. type
can be given as an integer in the range 1 to 255 inclusive,
designating objects with that internal typecode, or as a form
evaluating to a Lisp object, in which case that object's type is
designated. type should not be specified as a
Common Lisp type name, as that would be interpreted as designating
lists or symbols. (While typecodes are not programmatically available,
some are displayed in the output of
The optional size argument specifies the initial size of the hash-table used to record existing objects. In a large application or when tracking very common object types, it can be helpful to specify a larger initial size.
One global gc is performed before the body code is evaluated, to minimize unnecessary entries in the hash-table. Several are done after the body code is evaluated, to eliminate as much transient data as possible from the set of returned objects. Thus, the newly created vector in the body of this form:
(gcpath:collected-newstuff () (make-array 10))
will not appear in the returned result vector because it will be gc'ed away before the result is constructed while it will appear in the result of this form:
(gcpath:collected-newstuff () (setq my-vec (make-array 10)))
because it will still be the value of
the form completes.
The result is very close to the newly-allocated non-transient data created while the form is executed.
In a multiprocessing environment all new data allocated by all executing threads is considered, not just the data allocated by the thread executing the collected-newstuff form.
Copyright (c) 1998-2019, Franz Inc. Oakland, CA., USA. All rights reserved.
This page was not revised from the 10.0 page.
|Allegro CL version 10.1|
Unrevised from 10.0 to 10.1.