stan/jdk-24
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

8278824: Uneven work distribution when scanning heap roots in G1

Browse Source 
Reviewed-by: ayang, tschatzl, phh
This commit is contained in:
William Kemper 2021-12-28 17:32:00 +00:00 committed by Paul Hohensee
parent 54b800d56d
commit b4b0328d62
1 changed files with 2 additions and 2 deletions

4
src/hotspot/share/gc/g1/g1RemSet.cpp

@ -106,7 +106,7 @@ class G1RemSetScanState : public CHeapObj<mtGC> {
// within a region to claim. Dependent on the region size as proxy for the heap
// size, we limit the total number of chunks to limit memory usage and maintenance
// effort of that table vs. granularity of distributing scanning work.
// Testing showed that 8 for 1M/2M region, 16 for 4M/8M regions, 32 for 16/32M regions,
// Testing showed that 64 for 1M/2M region, 128 for 4M/8M regions, 256 for 16/32M regions,
// and so on seems to be such a good trade-off.
static uint get_chunks_per_region(uint log_region_size) {
// Limit the expected input values to current known possible values of the
@ -114,7 +114,7 @@ class G1RemSetScanState : public CHeapObj<mtGC> {
// values for region size.
assert(log_region_size >= 20 && log_region_size <= 29,
"expected value in [20,29], but got %u", log_region_size);
return 1u << (log_region_size / 2 - 7);
return 1u << (log_region_size / 2 - 4);
}
uint _scan_chunks_per_region; // Number of chunks per region.