On my arm64 server with 128 cores, 2 numa nodes.
I used memhog as benchmark :
numactl -m -C 5 memhog -r100000 1G
The test result as below:
With this patch:
#time migratepages 8490 0 1
real 0m1.161s
user 0m0.000s
sys 0m1.161s
without this patch:
#time migratepages 8460 0 1
real 0m2.068s
user 0m0.001s
sys 0m2.068s
So you can see the migration performance improvement about +78%
This is the perf record info.
w/o
+ 51.07% 0.09% migratepages [kernel.kallsyms] [k]
migrate_folio_extra
+ 42.43% 0.04% migratepages [kernel.kallsyms] [k]
folio_copy
+ 42.34% 42.34% migratepages [kernel.kallsyms] [k]
__pi_copy_page
+ 33.99% 0.09% migratepages [kernel.kallsyms] [k]
rmap_walk_anon
+ 32.35% 0.04% migratepages [kernel.kallsyms] [k]
try_to_migrate
+ 27.78% 27.78% migratepages [kernel.kallsyms] [k]
ptep_clear_flush
+ 8.19% 6.64% migratepages [kernel.kallsyms] [k]
folio_migrate_flagsmigrati_tlb_flush
w/ this patch
+ 18.57% 0.13% migratepages [kernel.kallsyms] [k]
migrate_pages
+ 18.23% 0.07% migratepages [kernel.kallsyms] [k]
migrate_pages_batch
+ 16.29% 0.13% migratepages [kernel.kallsyms] [k]
migrate_folio_move
+ 12.73% 0.10% migratepages [kernel.kallsyms] [k]
move_to_new_folio
+ 12.52% 0.06% migratepages [kernel.kallsyms] [k]
migrate_folio_extra
Therefore, this patch helps improve performance in page migration
So, you can add Tested-by: Xin Hao <xhao@xxxxxxxxxxxxxxxxx>
From: "Huang, Ying" <ying.huang@xxxxxxxxx> Now, migrate_pages() migrate folios one by one, like the fake code as follows, for each folio unmap flush TLB copy restore map If multiple folios are passed to migrate_pages(), there are opportunities to batch the TLB flushing and copying. That is, we can change the code to something as follows, for each folio unmap for each folio flush TLB for each folio copy for each folio restore map The total number of TLB flushing IPI can be reduced considerably. And we may use some hardware accelerator such as DSA to accelerate the folio copying. So in this patch, we refactor the migrate_pages() implementation and implement the TLB flushing batching. Base on this, hardware accelerated folio copying can be implemented. If too many folios are passed to migrate_pages(), in the naive batched implementation, we may unmap too many folios at the same time. The possibility for a task to wait for the migrated folios to be mapped again increases. So the latency may be hurt. To deal with this issue, the max number of folios be unmapped in batch is restricted to no more than HPAGE_PMD_NR in the unit of page. That is, the influence is at the same level of THP migration. We use the following test to measure the performance impact of the patchset, On a 2-socket Intel server, - Run pmbench memory accessing benchmark - Run `migratepages` to migrate pages of pmbench between node 0 and node 1 back and forth. With the patch, the TLB flushing IPI reduces 99.1% during the test and the number of pages migrated successfully per second increases 291.7%. This patchset is based on v6.2-rc4. Changes: v4: - Fixed another bug about non-LRU folio migration. Thanks Hyeonggon! v3: - Rebased on v6.2-rc4 - Fixed a bug about non-LRU folio migration. Thanks Mike! - Fixed some comments. Thanks Baolin! - Collected reviewed-by. v2: - Rebased on v6.2-rc3 - Fixed type force cast warning. Thanks Kees! - Added more comments and cleaned up the code. Thanks Andrew, Zi, Alistair, Dan! - Collected reviewed-by. from rfc to v1: - Rebased on v6.2-rc1 - Fix the deadlock issue caused by locking multiple pages synchronously per Alistair's comments. Thanks! - Fix the autonumabench panic per Rao's comments and fix. Thanks! - Other minor fixes per comments. Thanks! Best Regards, Huang, Ying
