Re: [RFC v5 00/11] Speculative page faults
From: Laurent Dufour
Date: Mon Jul 03 2017 - 13:33:15 EST
On 16/06/2017 19:52, Laurent Dufour wrote:
> This is a port on kernel 4.12 of the work done by Peter Zijlstra to
> handle page fault without holding the mm semaphore [1].
>
> The idea is to try to handle user space page faults without holding the
> mmap_sem. This should allow better concurrency for massively threaded
> process since the page fault handler will not wait for other threads memory
> layout change to be done, assuming that this change is done in another part
> of the process's memory space. This type page fault is named speculative
> page fault. If the speculative page fault fails because of a concurrency is
> detected or because underlying PMD or PTE tables are not yet allocating, it
> is failing its processing and a classic page fault is then tried.
>
> The speculative page fault (SPF) has to look for the VMA matching the fault
> address without holding the mmap_sem, so the VMA list is now managed using
> SRCU allowing lockless walking. The only impact would be the deferred file
> derefencing in the case of a file mapping, since the file pointer is
> released once the SRCU cleaning is done. This patch relies on the change
> done recently by Paul McKenney in SRCU which now runs a callback per CPU
> instead of per SRCU structure [1].
>
> The VMA's attributes checked during the speculative page fault processing
> have to be protected against parallel changes. This is done by using a per
> VMA sequence lock. This sequence lock allows the speculative page fault
> handler to fast check for parallel changes in progress and to abort the
> speculative page fault in that case.
>
> Once the VMA is found, the speculative page fault handler would check for
> the VMA's attributes to verify that the page fault has to be handled
> correctly or not. Thus the VMA is protected through a sequence lock which
> allows fast detection of concurrent VMA changes. If such a change is
> detected, the speculative page fault is aborted and a *classic* page fault
> is tried. VMA sequence lockings are added when VMA attributes which are
> checked during the page fault are modified.
>
> When the PTE is fetched, the VMA is checked to see if it has been changed,
> so once the page table is locked, the VMA is valid, so any other changes
> leading to touching this PTE will need to lock the page table, so no
> parallel change is possible at this time.
> .
> Compared to the Peter's initial work, this series introduces a spin_trylock
> when dealing with speculative page fault. This is required to avoid dead
> lock when handling a page fault while a TLB invalidate is requested by an
> other CPU holding the PTE. Another change due to a lock dependency issue
> with mapping->i_mmap_rwsem.
>
> This series builds on top of v4.12-rc5 and is functional on x86 and
> PowerPC.
>
> Tests have been made using a large commercial in-memory database on a
> PowerPC system with 752 CPUs. The results are very encouraging since the
> loading of the 2TB database was faster by 14% with the speculative page
> fault.
>
> However tests done using multi-fault [3] or kernbench [4], on smaller
> systems didn't show performance improvements, I saw a little degradation but
> running the tests again shows that this is in the noise. So nothing
> significant enough on the both sides.
>
> Since benchmarks are encouraging and running test suites didn't raise any
> issue, I'd like this request for comment series to move to a patch series
> soon. So please comment.
Hi all,
I didn't get any feedback for the moment on this series...
In addition to the previous benchmark results, here are some additional
metrics I captured using the ebizzy benchmark [A]. It is stressing the
mmap/munmap calls in an interesting way.
The test is counting the number of records per second it can manage, the
higher is the best. I run it like this 'ebizzy -mTRp'. To get consistent
result I repeat the test 100 times and measure the average result, mean
deviation and max. I run the test on top of 4.12 on 2 nodes, one with 80
CPUs, and the other one with 1024 CPUs:
* 80 CPUs Power 8 node:
Records/s 4.12 4.12-SPF
Average 38941,62 64235,82
Mean deviation 620,93 1718,95
Max 41988 69623
* 1024 CPUs Power 8 node:
Records/s 4.12 4.12-SPF
Average 39516,64 80689,27
Mean deviation 1387,66 1319,98
Max 43281 90441
So with that patch series, we got about 2x increase and the gap is growing
when the number of CPUs is higher.
Please comment as I'm about to drop the final patch rebased on 4.12.
Thanks,
Laurent.
[A] http://ebizzy.sourceforge.net/
> ------------------
> Benchmarks results
>
> Here are the results on a 8 CPUs X86 guest using kernbench on a 4.12-r5
> kernel (kernel is build 5 times):
>
> Average Half load -j 4 Run (std deviation):
> 4.12.0-rc5 4.12.0-rc5-spf
> Run (std deviation)
> Elapsed Time 48.42 (0.334515) 48.638 (0.344848)
> User Time 124.322 (0.964324) 124.478 (0.659902)
> System Time 58.008 (0.300865) 58.664 (0.590999)
> Percent CPU 376.2 (1.09545) 376.4 (1.51658)
> Context Switches 7409.6 (215.18) 11022.8 (281.093)
> Sleeps 15255.8 (63.0254) 15250.8 (43.4592)
>
> Average Optimal load -j 8
> 4.12.0-rc5 4.12.0-rc5-spf
> Run (std deviation)
> Elapsed Time 24.268 (0.151723) 24.514 (0.143805)
> User Time 112.092 (12.9135) 112.04 (13.1257)
> System Time 49.03 (9.46999) 49.721 (9.44455)
> Percent CPU 476 (105.205) 474.3 (103.209)
> Context Switches 10268.7 (3020.16) 14069.2 (3219.98)
> Sleeps 15790.8 (568.885) 15829.4 (615.371)
>
> Average Maximal load -j
> 4.12.0-rc5 4.12.0-rc5-spf
> Run (std deviation)
> Elapsed Time 25.042 (0.237844) 25.216 (0.201941)
> User Time 110.19 (10.7245) 110.312 (10.8245)
> System Time 45.9113 (8.86119) 46.48 (8.93778)
> Percent CPU 511.533 (99.1376) 510.133 (97.9897)
> Context Switches 19521.1 (13759.8) 22354.1 (12400)
> Sleeps 15514.7 (609.76) 15521.2 (670.054)
>
> The elapsed time is in the same order, a bit larger in the case of the spf
> release, but that seems to be in the error margin.
>
> Here are the kerbench results on a 572 CPUs Power8 system :
>
> Average Half load -j 376
> 4.12.0-rc5 4.12.0-rc5-spf
> Run (std deviation)
> Elapsed Time 3.384 (0.0680441) 3.344 (0.0634823)
> User Time 203.998 (8.41125) 193.476 (8.23406)
> System Time 13.064 (0.624444) 12.028 (0.495954)
> Percent CPU 6407 (285.422) 6136.2 (198.173)
> Context Switches 7319.2 (517.785) 8960 (221.735)
> Sleeps 24287.8 (861.132) 22902.4 (728.475)
>
> Average Optimal load -j 752
> 4.12.0-rc5 4.12.0-rc5-spf
> Run (std deviation)
> Elapsed Time 3.414 (0.136858) 3.432 (0.0506952)
> User Time 200.985 (8.71172) 197.747 (8.9511)
> System Time 12.903 (0.638262) 12.472 (0.684865)
> Percent CPU 6287.9 (322.208) 6194.8 (192.116)
> Context Switches 7173.5 (479.038) 9355.7 (712.3)
> Sleeps 24241.6 (1003.66) 22867.5 (1242.49)
>
> Average Maximal load -j
> 4.12.0-rc5 4.12.0-rc5-spf
> Run (std deviation)
> Elapsed Time 3.422 (0.0791833) 3.312 (0.109864)
> User Time 202.096 (7.45845) 197.541 (9.42758)
> System Time 12.8733 (0.57327) 12.4567 (0.568465)
> Percent CPU 6304.87 (278.195) 6234.67 (204.769)
> Context Switches 7166 (412.524) 9398.73 (639.917)
> Sleeps 24065.6 (1132.3) 22822.8 (1176.71)
>
> Here the elapsed time is a bit shorter using the spf release, but again we
> stay in the error margin.
>
> Here are results using multi-fault :
>
> --- x86 8 CPUs
> Page faults in 60s
> 4.12.0-rc5 23,014,776
> 4.12-0-rc5-spf 23,224,435
>
> --- ppc64le 752 CPUs
> Page faults in 60s
> 4.12.0-rc5 28,087,752
> 4.12-0-rc5-spf 32,272,610
>
> Results is a bit higher on ppc64le with the SPF patch, but I'm not convince
> about this test on Power8 since the page table are managed differently on
> this architecture, I'm wondering if we are not hitting the PTE lock.
> I run the test multiple times, the number are varying a bit but remain in
> the same order.
>
> ------------------
> Changes since V4:
> - merge several patches to reduce the series as requested by Jan Kara
> - check any comment warning in the code and remove each of them
> - reword some patch description
>
> Changes since V3:
> - support for the 5-level paging.
> - abort speculative path before entering userfault code
> - support for PowerPC architecture
> - reorder the patch to fix build test errors.
>
> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
> [2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=da915ad5cf25b5f5d358dd3670c3378d8ae8c03e
> [3] https://lkml.org/lkml/2010/1/6/28
> [4] http://ck.kolivas.org/apps/kernbench/kernbench-0.50/
>
> Laurent Dufour (5):
> mm: Introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
> mm: fix lock dependency against mapping->i_mmap_rwsem
> mm: Protect VMA modifications using VMA sequence count
> mm: Try spin lock in speculative path
> powerpc/mm: Add speculative page fault
>
> Peter Zijlstra (6):
> mm: Dont assume page-table invariance during faults
> mm: Prepare for FAULT_FLAG_SPECULATIVE
> mm: VMA sequence count
> mm: RCU free VMAs
> mm: Provide speculative fault infrastructure
> x86/mm: Add speculative pagefault handling
>
> arch/powerpc/mm/fault.c | 25 ++++-
> arch/x86/mm/fault.c | 14 +++
> fs/proc/task_mmu.c | 2 +
> include/linux/mm.h | 4 +
> include/linux/mm_types.h | 3 +
> kernel/fork.c | 1 +
> mm/init-mm.c | 1 +
> mm/internal.h | 20 ++++
> mm/madvise.c | 4 +
> mm/memory.c | 286 +++++++++++++++++++++++++++++++++++++++--------
> mm/mempolicy.c | 10 +-
> mm/mlock.c | 9 +-
> mm/mmap.c | 123 +++++++++++++++-----
> mm/mprotect.c | 2 +
> mm/mremap.c | 7 ++
> 15 files changed, 430 insertions(+), 81 deletions(-)
>