[Git][ghc/ghc][wip/T23397] rts: use live words to estimate heap size
Teo Camarasu (@teo)
gitlab at gitlab.haskell.org
Thu Feb 1 13:37:17 UTC 2024
Teo Camarasu pushed to branch wip/T23397 at Glasgow Haskell Compiler / GHC
Commits:
35ef8dc8 by Teo Camarasu at 2024-02-01T13:37:08+00:00
rts: use live words to estimate heap size
We use live words rather than live blocks to determine the size of the
heap for determining memory retention.
Most of the time these two metrics align, but they can come apart in
normal usage when using the nonmoving collector.
The nonmoving collector leads to a lot of partially occupied blocks. So,
using live words is more accurate.
They can also come apart when the heap is suffering from high levels
fragmentation caused by small pinned objects, but in this case, the
block size is the more accurate metric. Since this case is best avoided
anyway. It is ok to accept the trade-off that we might try (and
probably) fail to return more memory in this case.
See also the Note [Statistics for retaining memory]
Resolves #23397
- - - - -
4 changed files:
- docs/users_guide/9.10.1-notes.rst
- rts/sm/GC.c
- rts/sm/Storage.c
- rts/sm/Storage.h
Changes:
=====================================
docs/users_guide/9.10.1-notes.rst
=====================================
@@ -140,6 +140,11 @@ Runtime system
See :ghc-ticket:`23340`.
:rts-flag:`--nonmoving-dense-allocator-count=⟨count⟩` has been added to fine-tune this behaviour.
+- Memory return logic now uses live bytes rather than live blocks to measure the size of the heap.
+ This primarily affects the non-moving GC, which should now be more willing to return memory to the OS.
+ Users who have fine-tuned the :rts-flag:`-F ⟨factor⟩`, :rts-flag:`-Fd ⟨factor⟩`, or :rts-flag:`-O ⟨size⟩` flags,
+ and use the non-moving GC, should see if adjustments are needed in light of this change.
+
- Add a :rts-flag:`--no-automatic-time-samples` flag which stops time profiling samples being automatically started on
startup. Time profiling can be controlled manually using functions in ``GHC.Profiling``.
=====================================
rts/sm/GC.c
=====================================
@@ -1002,10 +1002,15 @@ GarbageCollect (struct GcConfig config,
need_copied_live = 0;
need_uncopied_live = 0;
for (i = 0; i < RtsFlags.GcFlags.generations; i++) {
- need_copied_live += genLiveCopiedBlocks(&generations[i]);
- need_uncopied_live += genLiveUncopiedBlocks(&generations[i]);
+ need_copied_live += genLiveCopiedWords(&generations[i]);
+ need_uncopied_live += genLiveUncopiedWords(&generations[i]);
}
+ // Convert the live words into live blocks
+ // See Note [Statistics for retaining memory]
+ need_copied_live = BLOCK_ROUND_UP(need_copied_live) / BLOCK_SIZE_W;
+ need_uncopied_live = BLOCK_ROUND_UP(need_uncopied_live) / BLOCK_SIZE_W;
+
debugTrace(DEBUG_gc, "(before) copied_live: %d; uncopied_live: %d", need_copied_live, need_uncopied_live );
@@ -1029,7 +1034,7 @@ GarbageCollect (struct GcConfig config,
ASSERT(need_uncopied_live + need_copied_live >= RtsFlags.GcFlags.minOldGenSize );
- debugTrace(DEBUG_gc, "(after) copyied_live: %d; uncopied_live: %d", need_copied_live, need_uncopied_live );
+ debugTrace(DEBUG_gc, "(after) copied_live: %d; uncopied_live: %d", need_copied_live, need_uncopied_live );
need_prealloc = 0;
for (i = 0; i < n_nurseries; i++) {
@@ -1067,7 +1072,7 @@ GarbageCollect (struct GcConfig config,
W_ scaled_needed = ((scaled_factor + unavoidable_copied_need_factor) * need_copied_live)
+ ((scaled_factor + unavoidable_uncopied_need_factor) * need_uncopied_live);
- debugTrace(DEBUG_gc, "factors_2: %f %d", ((scaled_factor + unavoidable_copied_need_factor) * need_copied_live), ((scaled_factor + unavoidable_uncopied_need_factor) * need_uncopied_live));
+ debugTrace(DEBUG_gc, "factors_2: %f %f", ((scaled_factor + unavoidable_copied_need_factor) * need_copied_live), ((scaled_factor + unavoidable_uncopied_need_factor) * need_uncopied_live));
need = need_prealloc + scaled_needed;
/* Also, if user set heap size, do not drop below it.
@@ -2422,3 +2427,39 @@ bool doIdleGCWork(Capability *cap STG_UNUSED, bool all)
*
*/
+
+/* Note [Statistics for retaining memory]
+* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+*
+* At the end of GC, we want to determine the size of the heap in order to
+* determine the amount of memory we wish to return to the OS, or if we want
+* to increase the heap size to the minimum.
+*
+* There's two promising candidates for this metric: live words, and live blocks.
+*
+* Measuring live blocks is promising because blocks are the smallest unit
+* that the storage manager can (de)allocate.
+* Most of the time live words and live blocks are very similar.
+*
+* But the two metrics can come apart when the heap is dominated
+* by small pinned objects, or when using the non-moving collector.
+*
+* In both cases, this happens because objects cannot be copied, so
+* block occupancy can fall as objects in a block become garbage.
+* In situations like this, using live blocks to determine memory
+* retention behaviour can lead to us being overly conservative.
+*
+* Instead we use live words rounded up to the block size to measure
+* heap size. This gives us a more accurate picture of the heap.
+*
+* This works particularly well with the nonmoving collector as we
+* can reuse the space taken up by dead heap objects. This choice is less good
+* for fragmentation caused by a few pinned objects retaining blocks.
+* In that case, the block can only be reused if it is deallocated in its entirety.
+* And therefore using live blocks would be more accurate in this case.
+* We assume that this is relatively rare and when it does happen,
+* this fragmentation is a problem that should be addressed in its own right.
+*
+* See: #23397
+*
+*/
=====================================
rts/sm/Storage.c
=====================================
@@ -1642,11 +1642,36 @@ W_ countOccupied (bdescr *bd)
return words;
}
-// Returns the total number of live blocks
+// Returns the total number of live words
W_ genLiveWords (generation *gen)
{
- return (gen->live_estimate ? gen->live_estimate : gen->n_words) +
- gen->n_large_words + gen->n_compact_blocks * BLOCK_SIZE_W;
+ return genLiveCopiedWords(gen) + genLiveUncopiedWords(gen);
+}
+
+// The number of live words which will be copied by the copying collector.
+W_ genLiveCopiedWords (generation *gen)
+{
+ if (gen == oldest_gen && RtsFlags.GcFlags.useNonmoving){
+ // the non-moving generation doesn't contain any copied data
+ return 0;
+ } else {
+ return gen->live_estimate ? gen->live_estimate : gen->n_words;
+ }
+}
+
+// The number of live words which will not be copied by the copying collector
+// This includes data living in non-moving collector segments, compact blocks and large/pinned blocks.
+W_ genLiveUncopiedWords(generation *gen)
+{
+ W_ nonmoving_blocks = 0;
+ // The nonmoving heap contains some blocks that live outside the regular generation structure.
+ if (gen == oldest_gen && RtsFlags.GcFlags.useNonmoving){
+ nonmoving_blocks =
+ (gen->live_estimate ? gen->live_estimate : gen->n_words)
+ + nonmoving_large_words
+ + nonmoving_compact_words;
+ }
+ return gen->n_large_words + gen->n_compact_blocks * BLOCK_SIZE_W + nonmoving_blocks;
}
// The number of live blocks which will be copied by the copying collector.
=====================================
rts/sm/Storage.h
=====================================
@@ -111,6 +111,8 @@ StgWord gcThreadLiveWords (uint32_t i, uint32_t g);
StgWord gcThreadLiveBlocks (uint32_t i, uint32_t g);
StgWord genLiveWords (generation *gen);
+StgWord genLiveCopiedWords (generation *gen);
+StgWord genLiveUncopiedWords (generation *gen);
StgWord genLiveBlocks (generation *gen);
StgWord genLiveCopiedBlocks (generation *gen);
StgWord genLiveUncopiedBlocks (generation *gen);
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/35ef8dc83428f5405e092b12eb8cfc440b6504d8
--
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/35ef8dc83428f5405e092b12eb8cfc440b6504d8
You're receiving this email because of your account on gitlab.haskell.org.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20240201/5d6bc94e/attachment-0001.html>
More information about the ghc-commits
mailing list