arch/arm64/include/asm/tlbflush.h | 18 ++ arch/x86/include/asm/tlbflush.h | 18 ++ arch/x86/mm/tlb.c | 2 - include/linux/mm.h | 22 ++ include/linux/mm_types.h | 39 ++- include/linux/sched.h | 10 + mm/compaction.c | 10 + mm/internal.h | 85 +++++- mm/memory.c | 8 + mm/migrate.c | 487 ++++++++++++++++++++++++++---- mm/page_alloc.c | 155 ++++++++-- mm/page_isolation.c | 6 + mm/page_reporting.c | 10 + mm/rmap.c | 40 ++- mm/swap.c | 20 +- 15 files changed, 826 insertions(+), 104 deletions(-)
Hi everyone,
While I'm working with a tiered memory system e.g. CXL memory, I have
been facing migration overhead esp. tlb shootdown on promotion or
demotion between different tiers. Yeah.. most tlb shootdowns on
migration through hinting fault can be avoided thanks to Huang Ying's
work, commit 4d4b6d66db ("mm,unmap: avoid flushing tlb in batch if PTE
is inaccessible"). See the following link for more information:
https://lore.kernel.org/lkml/20231115025755.GA29979@system.software.com/
However, it's only for ones using hinting fault. I thought it'd be much
better if we have a general mechanism to reduce all tlb numbers that we
can ultimately apply to any type of migration.
I'm suggesting a mechanism called MIGRC that stands for 'Migration Read
Copy', to reduce tlb numbers by deferring tlb flush until the source
folios at migration actually become used, of course, only if the target
PTE don't have write permission.
To achieve that:
1. For the folios that map only to non-writable tlb entries, prevent
tlb flush during migration but perform it just before the source
folios actually become used out of buddy or pcp.
2. When any non-writable tlb entry changes to writable e.g. through
fault handler, give up migrc mechanism and perform tlb flush
required right away.
No matter what type of workload is used for performance evaluation, the
result would be positive thanks to the unconditional reduction of tlb
flushes, tlb misses and interrupts. For the test, I picked up XSBench
that is widely used for performance analysis on high performance
computing architectures - https://github.com/ANL-CESAR/XSBench.
The result would depend on memory latency and how often reclaim runs,
which implies tlb miss overhead and how many times migration happens.
The slower the memory is and the more reclaim runs, the better migrc
works so as to obtain the better result. In my system, the result
shows:
1. itlb flushes are reduced over 90%.
2. itlb misses are reduced over 30%.
3. All the other tlb numbers also get enhanced.
4. tlb shootdown interrupts are reduced over 90%.
5. The test program runtime is reduced over 5%.
The test envitonment:
Architecture - x86_64
QEMU - kvm enabled, host cpu
Numa - 2 nodes (16 CPUs 1GB, no CPUs 99GB)
Linux Kernel - v6.9-rc4, numa balancing tiering on, demotion enabled
< measurement: raw data - tlb and interrupt numbers >
$ perf stat -a \
-e itlb.itlb_flush \
-e tlb_flush.dtlb_thread \
-e tlb_flush.stlb_any \
-e dtlb-load-misses \
-e dtlb-store-misses \
-e itlb-load-misses \
XSBench -t 16 -p 50000000
$ grep "TLB shootdowns" /proc/interrupts
BEFORE
------
40417078 itlb.itlb_flush
234852566 tlb_flush.dtlb_thread
153192357 tlb_flush.stlb_any
119001107892 dTLB-load-misses
307921167 dTLB-store-misses
1355272118 iTLB-load-misses
TLB: 1364803 1303670 1333921 1349607
1356934 1354216 1332972 1342842
1350265 1316443 1355928 1360793
1298239 1326358 1343006 1340971
TLB shootdowns
AFTER
-----
3316495 itlb.itlb_flush
138912511 tlb_flush.dtlb_thread
115199341 tlb_flush.stlb_any
117610390021 dTLB-load-misses
198042233 dTLB-store-misses
840066984 iTLB-load-misses
TLB: 117257 119219 117178 115737
117967 118948 117508 116079
116962 117266 117320 117215
105808 103934 115672 117610
TLB shootdowns
< measurement: user experience - runtime >
$ time XSBench -t 16 -p 50000000
BEFORE
------
Threads: 16
Runtime: 968.783 seconds
Lookups: 1,700,000,000
Lookups/s: 1,754,778
15208.91s user 141.44s system 1564% cpu 16:20.98 total
AFTER
-----
Threads: 16
Runtime: 913.210 seconds
Lookups: 1,700,000,000
Lookups/s: 1,861,565
14351.69s user 138.23s system 1565% cpu 15:25.47 total
---
Changes from v8:
1. Rebase on akpm/mm.git mm-unstable as of April 18, 2024.
2. Supplement comments and commit message.
3. Change the candidate to apply migrc mechnism:
BEFORE - The source folios at demotion and promotion.
AFTER - The souce folios at any type of migration.
4. Change how migrc mechanism works:
BEFORE - Reduce tlb flushes by deferring folio_free() for
source folios during demotion and promotion.
AFTER - Reduce tlb flushes by deferring tlb flush until they
actually become used, out of pcp or buddy. The
current version of migrc does *not* defer calling
folio_free() but let it go as it is as the same as
vanilla kernel, with the folios marked kind of 'need
to tlb flush'. And then handle the flush when the
page exits from pcp or buddy so as to prevent
changing vm stats e.g. free pages.
Changes from v7:
1. Rewrite cover letter to explain what 'migrc' mechasism is.
(feedbacked by Andrew Morton)
2. Supplement the commit message of a patch 'mm: Add APIs to
free a folio directly to the buddy bypassing pcp'.
(feedbacked by Andrew Morton)
Changes from v6:
1. Fix build errors in case of
CONFIG_ARCH_WANT_BATCHED_UNMAP_tlb_FLUSH disabled by moving
migrc_flush_{start,end}() calls from arch code to
try_to_unmap_flush() in mm/rmap.c.
Changes from v5:
1. Fix build errors in case of CONFIG_MIGRATION disabled or
CONFIG_HWPOISON_INJECT moduled. (feedbacked by kernel test
bot and Raymond Jay Golo)
2. Organize migrc code with two kconfigs, CONFIG_MIGRATION and
CONFIG_ARCH_WANT_BATCHED_UNMAP_tlb_FLUSH.
Changes from v4:
1. Rebase on v6.7.
2. Fix build errors in arm64 that is doing nothing for tlb flush
but has CONFIG_ARCH_WANT_BATCHED_UNMAP_tlb_FLUSH. (reported
by kernel test robot)
3. Don't use any page flag. So the system would give up migrc
mechanism more often but it's okay. The final improvement is
good enough.
4. Instead, optimize full tlb flush(arch_tlbbatch_flush()) by
avoiding redundant CPUs from tlb flush.
Changes from v3:
1. Don't use the kconfig, CONFIG_MIGRC, and remove sysctl knob,
migrc_enable. (feedbacked by Nadav)
2. Remove the optimization skipping CPUs that have already
performed tlb flushes needed by any reason when performing
tlb flushes by migrc because I can't tell the performance
difference between w/ the optimization and w/o that.
(feedbacked by Nadav)
3. Minimize arch-specific code. While at it, move all the migrc
declarations and inline functions from include/linux/mm.h to
mm/internal.h (feedbacked by Dave Hansen, Nadav)
4. Separate a part making migrc paused when the system is in
high memory pressure to another patch. (feedbacked by Nadav)
5. Rename:
a. arch_tlbbatch_clean() to arch_tlbbatch_clear(),
b. tlb_ubc_nowr to tlb_ubc_ro,
c. migrc_try_flush_free_folios() to migrc_flush_free_folios(),
d. migrc_stop to migrc_pause.
(feedbacked by Nadav)
6. Use ->lru list_head instead of introducing a new llist_head.
(feedbacked by Nadav)
7. Use non-atomic operations of page-flag when it's safe.
(feedbacked by Nadav)
8. Use stack instead of keeping a pointer of 'struct migrc_req'
in struct task, which is for manipulating it locally.
(feedbacked by Nadav)
9. Replace a lot of simple functions to inline functions placed
in a header, mm/internal.h. (feedbacked by Nadav)
10. Add additional sufficient comments. (feedbacked by Nadav)
11. Remove a lot of wrapper functions. (feedbacked by Nadav)
Changes from RFC v2:
1. Remove additional occupation in struct page. To do that,
unioned with lru field for migrc's list and added a page
flag. I know page flag is a thing that we don't like to add
but no choice because migrc should distinguish folios under
migrc's control from others. Instead, I force migrc to be
used only on 64 bit system to mitigate you guys from getting
angry.
2. Remove meaningless internal object allocator that I
introduced to minimize impact onto the system. However, a ton
of tests showed there was no difference.
3. Stop migrc from working when the system is in high memory
pressure like about to perform direct reclaim. At the
condition where the swap mechanism is heavily used, I found
the system suffered from regression without this control.
4. Exclude folios that pte_dirty() == true from migrc's interest
so that migrc can work simpler.
5. Combine several patches that work tightly coupled to one.
6. Add sufficient comments for better review.
7. Manage migrc's request in per-node manner (from globally).
8. Add tlb miss improvement in commit message.
9. Test with more CPUs(4 -> 16) to see bigger improvement.
Changes from RFC:
1. Fix a bug triggered when a destination folio at the previous
migration becomes a source folio at the next migration,
before the folio gets handled properly so that the folio can
play with another migration. There was inconsistency in the
folio's state. Fixed it.
2. Split the patch set into more pieces so that the folks can
review better. (Feedbacked by Nadav Amit)
3. Fix a wrong usage of barrier e.g. smp_mb__after_atomic().
(Feedbacked by Nadav Amit)
4. Tried to add sufficient comments to explain the patch set
better. (Feedbacked by Nadav Amit)
Byungchul Park (8):
x86/tlb: add APIs manipulating tlb batch's arch data
arm64: tlbflush: add APIs manipulating tlb batch's arch data
mm/rmap: recognize read-only tlb entries during batched tlb flush
x86/tlb, mm/rmap: separate arch_tlbbatch_clear() out of
arch_tlbbatch_flush()
mm: separate move/undo parts from migrate_pages_batch()
mm: buddy: make room for a new variable, mgen, in struct page
mm: add folio_put_mgen() to deliver migrc's generation number to pcp
or buddy
mm: defer tlb flush until the source folios at migration actually get
used
arch/arm64/include/asm/tlbflush.h | 18 ++
arch/x86/include/asm/tlbflush.h | 18 ++
arch/x86/mm/tlb.c | 2 -
include/linux/mm.h | 22 ++
include/linux/mm_types.h | 39 ++-
include/linux/sched.h | 10 +
mm/compaction.c | 10 +
mm/internal.h | 85 +++++-
mm/memory.c | 8 +
mm/migrate.c | 487 ++++++++++++++++++++++++++----
mm/page_alloc.c | 155 ++++++++--
mm/page_isolation.c | 6 +
mm/page_reporting.c | 10 +
mm/rmap.c | 40 ++-
mm/swap.c | 20 +-
15 files changed, 826 insertions(+), 104 deletions(-)
base-commit: f52bcd4a9f6058704a6f6b6b50418f579defd4fe
--
2.17.1
Byungchul Park <byungchul@sk.com> writes:
> Hi everyone,
>
> While I'm working with a tiered memory system e.g. CXL memory, I have
> been facing migration overhead esp. tlb shootdown on promotion or
> demotion between different tiers. Yeah.. most tlb shootdowns on
> migration through hinting fault can be avoided thanks to Huang Ying's
> work, commit 4d4b6d66db ("mm,unmap: avoid flushing tlb in batch if PTE
> is inaccessible"). See the following link for more information:
>
> https://lore.kernel.org/lkml/20231115025755.GA29979@system.software.com/
>
> However, it's only for ones using hinting fault. I thought it'd be much
> better if we have a general mechanism to reduce all tlb numbers that we
> can ultimately apply to any type of migration.
>
> I'm suggesting a mechanism called MIGRC that stands for 'Migration Read
> Copy', to reduce tlb numbers by deferring tlb flush until the source
> folios at migration actually become used, of course, only if the target
> PTE don't have write permission.
>
> To achieve that:
>
> 1. For the folios that map only to non-writable tlb entries, prevent
> tlb flush during migration but perform it just before the source
> folios actually become used out of buddy or pcp.
>
> 2. When any non-writable tlb entry changes to writable e.g. through
> fault handler, give up migrc mechanism and perform tlb flush
> required right away.
>
> No matter what type of workload is used for performance evaluation, the
> result would be positive thanks to the unconditional reduction of tlb
> flushes, tlb misses and interrupts. For the test, I picked up XSBench
> that is widely used for performance analysis on high performance
> computing architectures - https://github.com/ANL-CESAR/XSBench.
>
> The result would depend on memory latency and how often reclaim runs,
> which implies tlb miss overhead and how many times migration happens.
> The slower the memory is and the more reclaim runs, the better migrc
> works so as to obtain the better result. In my system, the result
> shows:
>
> 1. itlb flushes are reduced over 90%.
> 2. itlb misses are reduced over 30%.
> 3. All the other tlb numbers also get enhanced.
> 4. tlb shootdown interrupts are reduced over 90%.
> 5. The test program runtime is reduced over 5%.
>
> The test envitonment:
>
> Architecture - x86_64
> QEMU - kvm enabled, host cpu
The test is run in VM? Do you have test results in bare metal
environment?
> Numa - 2 nodes (16 CPUs 1GB, no CPUs 99GB)
The configuration looks quite abnormal. Have you tested with other
configuration, such 1:4 or 1:8?
> Linux Kernel - v6.9-rc4, numa balancing tiering on, demotion enabled
>
> < measurement: raw data - tlb and interrupt numbers >
>
> $ perf stat -a \
> -e itlb.itlb_flush \
> -e tlb_flush.dtlb_thread \
> -e tlb_flush.stlb_any \
> -e dtlb-load-misses \
> -e dtlb-store-misses \
> -e itlb-load-misses \
> XSBench -t 16 -p 50000000
>
> $ grep "TLB shootdowns" /proc/interrupts
>
> BEFORE
> ------
> 40417078 itlb.itlb_flush
> 234852566 tlb_flush.dtlb_thread
> 153192357 tlb_flush.stlb_any
> 119001107892 dTLB-load-misses
> 307921167 dTLB-store-misses
> 1355272118 iTLB-load-misses
>
> TLB: 1364803 1303670 1333921 1349607
> 1356934 1354216 1332972 1342842
> 1350265 1316443 1355928 1360793
> 1298239 1326358 1343006 1340971
> TLB shootdowns
>
> AFTER
> -----
> 3316495 itlb.itlb_flush
> 138912511 tlb_flush.dtlb_thread
> 115199341 tlb_flush.stlb_any
> 117610390021 dTLB-load-misses
> 198042233 dTLB-store-misses
> 840066984 iTLB-load-misses
>
> TLB: 117257 119219 117178 115737
> 117967 118948 117508 116079
> 116962 117266 117320 117215
> 105808 103934 115672 117610
> TLB shootdowns
>
> < measurement: user experience - runtime >
>
> $ time XSBench -t 16 -p 50000000
>
> BEFORE
> ------
> Threads: 16
> Runtime: 968.783 seconds
> Lookups: 1,700,000,000
> Lookups/s: 1,754,778
>
> 15208.91s user 141.44s system 1564% cpu 16:20.98 total
>
> AFTER
> -----
> Threads: 16
> Runtime: 913.210 seconds
> Lookups: 1,700,000,000
> Lookups/s: 1,861,565
>
> 14351.69s user 138.23s system 1565% cpu 15:25.47 total
IIUC, the memory footprint will be larger with the patchset. Do you
have data?
--
Best Regards,
Huang, Ying
On Fri, Apr 19, 2024 at 02:06:30PM +0800, Huang, Ying wrote:
> Byungchul Park <byungchul@sk.com> writes:
>
> > The test envitonment:
> >
> > Architecture - x86_64
> > QEMU - kvm enabled, host cpu
>
> The test is run in VM? Do you have test results in bare metal
> environment?
I tested it in a bare metal server. See the result below.
> > Numa - 2 nodes (16 CPUs 1GB, no CPUs 99GB)
>
> The configuration looks quite abnormal. Have you tested with other
> configuration, such 1:4 or 1:8?
I tested with DRAM : CXL expander = 42GB : 98GB.
> > Linux Kernel - v6.9-rc4, numa balancing tiering on, demotion enabled
> >
> > < measurement: raw data - tlb and interrupt numbers >
> >
> > $ perf stat -a \
> > -e itlb.itlb_flush \
> > -e tlb_flush.dtlb_thread \
> > -e tlb_flush.stlb_any \
> > -e dtlb-load-misses \
> > -e dtlb-store-misses \
> > -e itlb-load-misses \
> > XSBench -t 16 -p 50000000
> >
> > $ grep "TLB shootdowns" /proc/interrupts
> >
> > BEFORE
> > ------
> > 40417078 itlb.itlb_flush
> > 234852566 tlb_flush.dtlb_thread
> > 153192357 tlb_flush.stlb_any
> > 119001107892 dTLB-load-misses
> > 307921167 dTLB-store-misses
> > 1355272118 iTLB-load-misses
> >
> > TLB: 1364803 1303670 1333921 1349607
> > 1356934 1354216 1332972 1342842
> > 1350265 1316443 1355928 1360793
> > 1298239 1326358 1343006 1340971
> > TLB shootdowns
> >
> > AFTER
> > -----
> > 3316495 itlb.itlb_flush
> > 138912511 tlb_flush.dtlb_thread
> > 115199341 tlb_flush.stlb_any
> > 117610390021 dTLB-load-misses
> > 198042233 dTLB-store-misses
> > 840066984 iTLB-load-misses
> >
> > TLB: 117257 119219 117178 115737
> > 117967 118948 117508 116079
> > 116962 117266 117320 117215
> > 105808 103934 115672 117610
> > TLB shootdowns
> >
> > < measurement: user experience - runtime >
> >
> > $ time XSBench -t 16 -p 50000000
> >
> > BEFORE
> > ------
> > Threads: 16
> > Runtime: 968.783 seconds
> > Lookups: 1,700,000,000
> > Lookups/s: 1,754,778
> >
> > 15208.91s user 141.44s system 1564% cpu 16:20.98 total
> >
> > AFTER
> > -----
> > Threads: 16
> > Runtime: 913.210 seconds
> > Lookups: 1,700,000,000
> > Lookups/s: 1,861,565
> >
> > 14351.69s user 138.23s system 1565% cpu 15:25.47 total
>
> IIUC, the memory footprint will be larger with the patchset. Do you
> have data?
As I already told you, from version 9, the footprint is exactly same
between patched kernel and vanilla kernel because that let folios go as
is, but controls TLB flush timing only.
There's two things to note.
1. I changed the patchset and will post the next version shortly:
BEFORE - Defer TLB flush required until the interesting folios
exiting either pcp or buddy. The interesting folios
are source folios unmapped during folio migration.
AFTER - Defer TLB flush required until the interesting folios
exiting either pcp or buddy. The interesting folios
are source folios unmapped during folio migration,
* plus, folios unmapped during reclaiming folios in
shrink_folio_list()*.
2. I changed workload for testing because XSBench doesn't struggle
against lack of memory in such a big server. Instead, I picked a
very real workload, LLM inference engine, llama.cpp.
I tested with the two changes. The test result is like:
---
Kernel version: mm-unstable around v6.9-rc4
Machine: bare metal, x86_64, Intel(R) Xeon(R) Gold 6430
CPU: 1 socket 64 core with hyper thread on
Numa: 2 nodes (64 CPUs DRAM 42GB, no CPUs CXL(expander) 98GB)
Config: swap off, numa balancing tiering on, demotion enabled
1 set of test workload:
echo 3 > /proc/sys/vm/drop_caches
llama.cpp/main -m $(70G_model1) -p "who are you?" -s 1 -t 15 -n 20 &
llama.cpp/main -m $(70G_model2) -p "who are you?" -s 1 -t 15 -n 20 &
llama.cpp/main -m $(70G_model3) -p "who are you?" -s 1 -t 15 -n 20 &
wait
where -t: nr of threads, -s: seed used to make the runtime stable,
-n: nr of tokens determinig the runtime, -p: prompt to ask, -m: LLM
model to use.
Run this set 10 times successively. So I got 30 total runtimes since
each inference prints its runtime at the end of each run. The result
is like:
BEFORE
------
llama_print_timings: total time = 1002461.95 ms / 24 tokens
llama_print_timings: total time = 1044978.38 ms / 24 tokens
llama_print_timings: total time = 1000653.09 ms / 24 tokens
llama_print_timings: total time = 1047104.80 ms / 24 tokens
llama_print_timings: total time = 1069430.36 ms / 24 tokens
llama_print_timings: total time = 1068201.16 ms / 24 tokens
llama_print_timings: total time = 1078092.59 ms / 24 tokens
llama_print_timings: total time = 1073200.45 ms / 24 tokens
llama_print_timings: total time = 1067136.00 ms / 24 tokens
llama_print_timings: total time = 1076442.56 ms / 24 tokens
llama_print_timings: total time = 1004142.64 ms / 24 tokens
llama_print_timings: total time = 1042942.65 ms / 24 tokens
llama_print_timings: total time = 999933.76 ms / 24 tokens
llama_print_timings: total time = 1046548.83 ms / 24 tokens
llama_print_timings: total time = 1068671.48 ms / 24 tokens
llama_print_timings: total time = 1068285.76 ms / 24 tokens
llama_print_timings: total time = 1077789.63 ms / 24 tokens
llama_print_timings: total time = 1071558.93 ms / 24 tokens
llama_print_timings: total time = 1066181.55 ms / 24 tokens
llama_print_timings: total time = 1076767.53 ms / 24 tokens
llama_print_timings: total time = 1004065.63 ms / 24 tokens
llama_print_timings: total time = 1044522.13 ms / 24 tokens
llama_print_timings: total time = 999725.33 ms / 24 tokens
llama_print_timings: total time = 1047510.77 ms / 24 tokens
llama_print_timings: total time = 1068010.27 ms / 24 tokens
llama_print_timings: total time = 1068999.31 ms / 24 tokens
llama_print_timings: total time = 1077648.05 ms / 24 tokens
llama_print_timings: total time = 1071378.96 ms / 24 tokens
llama_print_timings: total time = 1066326.32 ms / 24 tokens
llama_print_timings: total time = 1077088.92 ms / 24 tokens
AFTER
-----
llama_print_timings: total time = 988522.03 ms / 24 tokens
llama_print_timings: total time = 997204.52 ms / 24 tokens
llama_print_timings: total time = 996605.86 ms / 24 tokens
llama_print_timings: total time = 991985.50 ms / 24 tokens
llama_print_timings: total time = 1035143.31 ms / 24 tokens
llama_print_timings: total time = 993660.18 ms / 24 tokens
llama_print_timings: total time = 983082.14 ms / 24 tokens
llama_print_timings: total time = 990431.36 ms / 24 tokens
llama_print_timings: total time = 992707.09 ms / 24 tokens
llama_print_timings: total time = 992673.27 ms / 24 tokens
llama_print_timings: total time = 989285.43 ms / 24 tokens
llama_print_timings: total time = 996710.06 ms / 24 tokens
llama_print_timings: total time = 996534.64 ms / 24 tokens
llama_print_timings: total time = 991344.17 ms / 24 tokens
llama_print_timings: total time = 1035210.84 ms / 24 tokens
llama_print_timings: total time = 994714.13 ms / 24 tokens
llama_print_timings: total time = 984184.15 ms / 24 tokens
llama_print_timings: total time = 990909.45 ms / 24 tokens
llama_print_timings: total time = 991881.48 ms / 24 tokens
llama_print_timings: total time = 993918.03 ms / 24 tokens
llama_print_timings: total time = 990061.34 ms / 24 tokens
llama_print_timings: total time = 998076.69 ms / 24 tokens
llama_print_timings: total time = 997082.59 ms / 24 tokens
llama_print_timings: total time = 990677.58 ms / 24 tokens
llama_print_timings: total time = 1036054.94 ms / 24 tokens
llama_print_timings: total time = 994125.93 ms / 24 tokens
llama_print_timings: total time = 982467.01 ms / 24 tokens
llama_print_timings: total time = 990191.60 ms / 24 tokens
llama_print_timings: total time = 993319.24 ms / 24 tokens
llama_print_timings: total time = 992540.57 ms / 24 tokens
The difference of TLB shootdown(/proc/interrupts) is like:
BEFORE
------
TLB:
125553646 141418810 161932620 176853972 186655697 190399283
192143823 196414038 192872439 193313658 193395617 192521416
190788161 195067598 198016061 193607347 194293972 190786732
191545637 194856822 191801931 189634535 190399803 196365922
195268398 190115840 188050050 193194908 195317617 190820190
190164820 185556071 226797214 229592631 216112464 209909495
205575979 205950252 204948111 197999795 198892232 205287952
199344631 195015158 195869844 198858745 195692876 200961904
203463252 205921722 199850838 206145986 199613202 199961345
200129577 203020521 207873649 203697671 197093386 204243803
205993323 200934664 204193128 194435376 TLB shootdowns
AFTER
-----
TLB:
5648092 6610142 7032849 7882308 8088518 8352310
8656536 8705136 8647426 8905583 8985408 8704522
8884344 9026261 8929974 8869066 8877575 8810096
8770984 8754503 8801694 8865925 8787524 8656432
8755912 8682034 8773935 8832925 8797997 8515777
8481240 8891258 10595243 10285973 9756935 9573681
9398968 9069244 9242984 8899009 9310690 9029095
9069758 9105825 9092703 9270202 9460287 9258546
9180415 9232723 9270611 9175020 9490420 9360316
9420818 9057663 9525631 9310152 9152242 8654483
9181804 9050847 8919916 8883856 TLB shootdowns
The difference of 'perf stat' for tlb numbers during one set of test
with drop cache excluded is like:
BEFORE
------
3163679332 dTLB-load-misses
2017751856 dTLB-store-misses
327092903 iTLB-load-misses
1357543886 tlb:tlb_flush
AFTER
-----
2394694609 dTLB-load-misses
861144167 dTLB-store-misses
64055579 iTLB-load-misses
69175002 tlb:tlb_flush
---
I'm happy to share great results. I used a real workload that is super
popular these days, and got good results. I will post the next version
of the patchset shortly after organizing and refining things.
Byungchul
On Fri, Apr 19, 2024 at 02:06:30PM +0800, Huang, Ying wrote:
> Byungchul Park <byungchul@sk.com> writes:
>
> > Hi everyone,
> >
> > While I'm working with a tiered memory system e.g. CXL memory, I have
> > been facing migration overhead esp. tlb shootdown on promotion or
> > demotion between different tiers. Yeah.. most tlb shootdowns on
> > migration through hinting fault can be avoided thanks to Huang Ying's
> > work, commit 4d4b6d66db ("mm,unmap: avoid flushing tlb in batch if PTE
> > is inaccessible"). See the following link for more information:
> >
> > https://lore.kernel.org/lkml/20231115025755.GA29979@system.software.com/
> >
> > However, it's only for ones using hinting fault. I thought it'd be much
> > better if we have a general mechanism to reduce all tlb numbers that we
> > can ultimately apply to any type of migration.
> >
> > I'm suggesting a mechanism called MIGRC that stands for 'Migration Read
> > Copy', to reduce tlb numbers by deferring tlb flush until the source
> > folios at migration actually become used, of course, only if the target
> > PTE don't have write permission.
> >
> > To achieve that:
> >
> > 1. For the folios that map only to non-writable tlb entries, prevent
> > tlb flush during migration but perform it just before the source
> > folios actually become used out of buddy or pcp.
> >
> > 2. When any non-writable tlb entry changes to writable e.g. through
> > fault handler, give up migrc mechanism and perform tlb flush
> > required right away.
> >
> > No matter what type of workload is used for performance evaluation, the
> > result would be positive thanks to the unconditional reduction of tlb
> > flushes, tlb misses and interrupts. For the test, I picked up XSBench
> > that is widely used for performance analysis on high performance
> > computing architectures - https://github.com/ANL-CESAR/XSBench.
> >
> > The result would depend on memory latency and how often reclaim runs,
> > which implies tlb miss overhead and how many times migration happens.
> > The slower the memory is and the more reclaim runs, the better migrc
> > works so as to obtain the better result. In my system, the result
> > shows:
> >
> > 1. itlb flushes are reduced over 90%.
> > 2. itlb misses are reduced over 30%.
> > 3. All the other tlb numbers also get enhanced.
> > 4. tlb shootdown interrupts are reduced over 90%.
> > 5. The test program runtime is reduced over 5%.
> >
> > The test envitonment:
> >
> > Architecture - x86_64
> > QEMU - kvm enabled, host cpu
>
> The test is run in VM? Do you have test results in bare metal
> environment?
I will test in a bare metal environment and share the result.
> > Numa - 2 nodes (16 CPUs 1GB, no CPUs 99GB)
>
> The configuration looks quite abnormal. Have you tested with other
> configuration, such 1:4 or 1:8?
Okay I will test with the configurations.
> > Linux Kernel - v6.9-rc4, numa balancing tiering on, demotion enabled
> >
> > < measurement: raw data - tlb and interrupt numbers >
> >
> > $ perf stat -a \
> > -e itlb.itlb_flush \
> > -e tlb_flush.dtlb_thread \
> > -e tlb_flush.stlb_any \
> > -e dtlb-load-misses \
> > -e dtlb-store-misses \
> > -e itlb-load-misses \
> > XSBench -t 16 -p 50000000
> >
> > $ grep "TLB shootdowns" /proc/interrupts
> >
> > BEFORE
> > ------
> > 40417078 itlb.itlb_flush
> > 234852566 tlb_flush.dtlb_thread
> > 153192357 tlb_flush.stlb_any
> > 119001107892 dTLB-load-misses
> > 307921167 dTLB-store-misses
> > 1355272118 iTLB-load-misses
> >
> > TLB: 1364803 1303670 1333921 1349607
> > 1356934 1354216 1332972 1342842
> > 1350265 1316443 1355928 1360793
> > 1298239 1326358 1343006 1340971
> > TLB shootdowns
> >
> > AFTER
> > -----
> > 3316495 itlb.itlb_flush
> > 138912511 tlb_flush.dtlb_thread
> > 115199341 tlb_flush.stlb_any
> > 117610390021 dTLB-load-misses
> > 198042233 dTLB-store-misses
> > 840066984 iTLB-load-misses
> >
> > TLB: 117257 119219 117178 115737
> > 117967 118948 117508 116079
> > 116962 117266 117320 117215
> > 105808 103934 115672 117610
> > TLB shootdowns
> >
> > < measurement: user experience - runtime >
> >
> > $ time XSBench -t 16 -p 50000000
> >
> > BEFORE
> > ------
> > Threads: 16
> > Runtime: 968.783 seconds
> > Lookups: 1,700,000,000
> > Lookups/s: 1,754,778
> >
> > 15208.91s user 141.44s system 1564% cpu 16:20.98 total
> >
> > AFTER
> > -----
> > Threads: 16
> > Runtime: 913.210 seconds
> > Lookups: 1,700,000,000
> > Lookups/s: 1,861,565
> >
> > 14351.69s user 138.23s system 1565% cpu 15:25.47 total
>
> IIUC, the memory footprint will be larger with the patchset. Do you
> have data?
No. The footprint is, I expect, same as vanilla with this patchset. I
will share the data.
Last time, since you pointed out that the footprint seemed to be larger
with the previous patchset becasue it worked anyway based on deferring
freeing folios.
Which made me rework on it so as to avoid tweaking the original behavior
of mm. Instead, the current version of migrc let it go exactly same as
it is with vanilla until the interesting folios exit from pcp or buddy,
and do tlb flush if needed.
Byungchul
> --
> Best Regards,
> Huang, Ying
On Thu, 18 Apr 2024 15:15:28 +0900 Byungchul Park <byungchul@sk.com> wrote: > $ time XSBench -t 16 -p 50000000 > > BEFORE > ------ > Threads: 16 > Runtime: 968.783 seconds > Lookups: 1,700,000,000 > Lookups/s: 1,754,778 > > 15208.91s user 141.44s system 1564% cpu 16:20.98 total > > AFTER > ----- > Threads: 16 > Runtime: 913.210 seconds > Lookups: 1,700,000,000 > Lookups/s: 1,861,565 > > 14351.69s user 138.23s system 1565% cpu 15:25.47 total Well that's nice. What exactly is XSBench doing in this situation? What sort of improvements can we expect to see in useful workloads? I see it no longer consumes an additional page flag, good. The patches show no evidence of review activity and I'm not seeing much on the mailing list (patchset title was changed. Previous title "Reduce TLB flushes under some specific conditions"). Perhaps a better description of the overall benefit to our users would help to motivate reviewers.
On Thu, Apr 18, 2024 at 01:17:57PM -0700, Andrew Morton wrote:
> On Thu, 18 Apr 2024 15:15:28 +0900 Byungchul Park <byungchul@sk.com> wrote:
>
> > $ time XSBench -t 16 -p 50000000
> >
> > BEFORE
> > ------
> > Threads: 16
> > Runtime: 968.783 seconds
> > Lookups: 1,700,000,000
> > Lookups/s: 1,754,778
> >
> > 15208.91s user 141.44s system 1564% cpu 16:20.98 total
> >
> > AFTER
> > -----
> > Threads: 16
> > Runtime: 913.210 seconds
> > Lookups: 1,700,000,000
> > Lookups/s: 1,861,565
> >
> > 14351.69s user 138.23s system 1565% cpu 15:25.47 total
>
> Well that's nice. What exactly is XSBench doing in this situation?
As far as I know, it's frequently and continuously accessing annon areas
with addresses ranged within 6GB, by multi threads. Thus, it triggers a
lot of promotions by hinting fault of numa balancing tiering and a lot
of demotions by kswapd as well, resulting in a ton of tlb flushes.
All I need is a system suffering from memory reclaim or any type of
folio migration since migrc mechanism is one for mitigating the overhead
of folio migration. To see the benefits of migrc, it doesn't have to be
XSBench but any workload suffering from reclaim.
> What sort of improvements can we expect to see in useful workloads?
Increase throughput(= runtime reduction of each work in the system).
1. Because migrc removes the CPU time that would've been spent in IPI
handler due to tlb shootdown, by skipping a lot of tlb shootdowns.
2. Becasue migrc reduces tlb misses so as to utilize tlb cache better,
by skipping a lof of tlb flushes.
Besides, I expect overall scheduler latencies can be enhanced, the worst
latencies measured using some tracters of ftrace showed no change though.
> I see it no longer consumes an additional page flag, good.
>
> The patches show no evidence of review activity and I'm not seeing much
> on the mailing list (patchset title was changed. Previous title
> "Reduce TLB flushes under some specific conditions"). Perhaps a better
I changed the title because it was supposed to work with only numa
balancing tiering like promotion and demotion but, for now, migrc works
with any type of folio migration. Thus, I can tell migrc demonstrates
its benefits as long as a system is under the control of reclaim and
folio migration.
Byungchul
> description of the overall benefit to our users would help to motivate
> reviewers.
© 2016 - 2026 Red Hat, Inc.