Re: [PATCH] tmpfs: avoid a little creat and stat slowdown
On Fri, 13 Nov 2015, Huang, Ying wrote: > > c435a390574d is the direct parent of afa2db2fb6f1 in its original git. > 43819159da2b is your patch applied on top of v4.3-rc7. The comparison > of 43819159da2b with v4.3-rc7 is as follow: ... > So you patch improved 11.9% from its base v4.3-rc7. I think other > difference are caused by other changes. Sorry for confusing. Thanks for getting back on this: that's rather what I was hoping to hear. Of course, no user will care which commit is responsible for a slowdown, and we may need to look further; but I couldn't make sense of it before, so this was a relief. Hugh -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] tmpfs: avoid a little creat and stat slowdown
Hugh Dickins writes: > On Wed, 4 Nov 2015, Huang, Ying wrote: >> Hugh Dickins writes: >> >> > LKP reports that v4.2 commit afa2db2fb6f1 ("tmpfs: truncate prealloc >> > blocks past i_size") causes a 14.5% slowdown in the AIM9 creat-clo >> > benchmark. >> > >> > creat-clo does just what you'd expect from the name, and creat's O_TRUNC >> > on 0-length file does indeed get into more overhead now shmem_setattr() >> > tests "0 <= 0" instead of "0 < 0". >> > >> > I'm not sure how much we care, but I think it would not be too VW-like >> > to add in a check for whether any pages (or swap) are allocated: if none >> > are allocated, there's none to remove from the radix_tree. At first I >> > thought that check would be good enough for the unmaps too, but no: we >> > should not skip the unlikely case of unmapping pages beyond the new EOF, >> > which were COWed from holes which have now been reclaimed, leaving none. >> > >> > This gives me an 8.5% speedup: on Haswell instead of LKP's Westmere, >> > and running a debug config before and after: I hope those account for >> > the lesser speedup. >> > >> > And probably someone has a benchmark where a thousand threads keep on >> > stat'ing the same file repeatedly: forestall that report by adjusting >> > v4.3 commit 44a30220bc0a ("shmem: recalculate file inode when fstat") >> > not to take the spinlock in shmem_getattr() when there's no work to do. >> > >> > Reported-by: Ying Huang >> > Signed-off-by: Hugh Dickins >> >> Hi, Hugh, >> >> Thanks a lot for your support! The test on LKP shows that this patch >> restores a big part of the regression! In following list, >> >> c435a390574d012f8d30074135d8fcc6f480b484: is parent commit >> afa2db2fb6f15f860069de94a1257db57589fe95: is the first bad commit has >> performance regression. >> 43819159da2b77fedcf7562134d6003dccd6a068: is the fixing patch > > Hi Ying, > > Thank you, for reporting, and for trying out the patch (which is now > in Linus's tree as commit d0424c429f8e0555a337d71e0a13f2289c636ec9). > > But I'm disappointed by the result: do I understand correctly, > that afa2db2fb6f1 made a -12.5% change, but the fix still -5.6% > from your parent comparison point? Yes. > If we value that microbenchmark > at all (debatable), I'd say that's not good enough. I think that is a good improvement. > It does match with my own rough measurement, but I'd been hoping > for better when done in a more controlled environment; and I cannot > explain why "truncate prealloc blocks past i_size" creat-clo performance > would not be fully corrected by "avoid a little creat and stat slowdown" > (unless either patch adds subtle icache or dcache displacements). > > I'm not certain of how you performed the comparison. Was the > c435a390574d tree measured, then patch afa2db2fb6f1 applied on top > of that and measured, then patch 43819159da2b applied on top of that > and measured? Or were there other intervening changes, which could > easily add their own interference? c435a390574d is the direct parent of afa2db2fb6f1 in its original git. 43819159da2b is your patch applied on top of v4.3-rc7. The comparison of 43819159da2b with v4.3-rc7 is as follow: = compiler/cpufreq_governor/kconfig/rootfs/tbox_group/test/testcase/testtime: gcc-4.9/performance/x86_64-rhel/debian-x86_64-2015-02-07.cgz/lkp-wsx02/creat-clo/aim9/300s commit: 32b88194f71d6ae7768a29f87fbba454728273ee 43819159da2b77fedcf7562134d6003dccd6a068 32b88194f71d6ae7 43819159da2b77fedcf7562134 -- %stddev %change %stddev \ |\ 475224 ± 1% +11.9% 531968 ± 1% aim9.creat-clo.ops_per_sec 10469094 ±201% -52.3%4998529 ±130% latency_stats.avg.nfs_wait_on_request.nfs_updatepage.nfs_write_end.generic_perform_write.__generic_file_write_iter.generic_file_write_iter.nfs_file_write.__vfs_write.vfs_write.SyS_write.entry_SYSCALL_64_fastpath 18852332 ±223% -73.5%4998529 ±130% latency_stats.max.nfs_wait_on_request.nfs_updatepage.nfs_write_end.generic_perform_write.__generic_file_write_iter.generic_file_write_iter.nfs_file_write.__vfs_write.vfs_write.SyS_write.entry_SYSCALL_64_fastpath 21758590 ±199% -77.0%4998529 ±130% latency_stats.sum.nfs_wait_on_request.nfs_updatepage.nfs_write_end.generic_perform_write.__generic_file_write_iter.generic_file_write_iter.nfs_file_write.__vfs_write.vfs_write.SyS_write.entry_SYSCALL_64_fastpath 4817724 ± 0% +9.6%5280303 ± 1% proc-vmstat.numa_hit 4812582 ± 0% +9.7%5280287 ± 1% proc-vmstat.numa_local 8499767 ± 4% +14.2%9707953 ± 4% proc-vmstat.pgalloc_normal 8984075 ± 0% +10.4%9919044 ± 1% proc-vmstat.pgfree 9.22 ± 8% +27.4% 11.75 ± 9% sched_debug.cfs_rq[0]:/.nr_spread_over 2667 ± 63% +90.0% 5068 ± 37%
Re: [PATCH] tmpfs: avoid a little creat and stat slowdown
Hugh Dickinswrites: > On Wed, 4 Nov 2015, Huang, Ying wrote: >> Hugh Dickins writes: >> >> > LKP reports that v4.2 commit afa2db2fb6f1 ("tmpfs: truncate prealloc >> > blocks past i_size") causes a 14.5% slowdown in the AIM9 creat-clo >> > benchmark. >> > >> > creat-clo does just what you'd expect from the name, and creat's O_TRUNC >> > on 0-length file does indeed get into more overhead now shmem_setattr() >> > tests "0 <= 0" instead of "0 < 0". >> > >> > I'm not sure how much we care, but I think it would not be too VW-like >> > to add in a check for whether any pages (or swap) are allocated: if none >> > are allocated, there's none to remove from the radix_tree. At first I >> > thought that check would be good enough for the unmaps too, but no: we >> > should not skip the unlikely case of unmapping pages beyond the new EOF, >> > which were COWed from holes which have now been reclaimed, leaving none. >> > >> > This gives me an 8.5% speedup: on Haswell instead of LKP's Westmere, >> > and running a debug config before and after: I hope those account for >> > the lesser speedup. >> > >> > And probably someone has a benchmark where a thousand threads keep on >> > stat'ing the same file repeatedly: forestall that report by adjusting >> > v4.3 commit 44a30220bc0a ("shmem: recalculate file inode when fstat") >> > not to take the spinlock in shmem_getattr() when there's no work to do. >> > >> > Reported-by: Ying Huang >> > Signed-off-by: Hugh Dickins >> >> Hi, Hugh, >> >> Thanks a lot for your support! The test on LKP shows that this patch >> restores a big part of the regression! In following list, >> >> c435a390574d012f8d30074135d8fcc6f480b484: is parent commit >> afa2db2fb6f15f860069de94a1257db57589fe95: is the first bad commit has >> performance regression. >> 43819159da2b77fedcf7562134d6003dccd6a068: is the fixing patch > > Hi Ying, > > Thank you, for reporting, and for trying out the patch (which is now > in Linus's tree as commit d0424c429f8e0555a337d71e0a13f2289c636ec9). > > But I'm disappointed by the result: do I understand correctly, > that afa2db2fb6f1 made a -12.5% change, but the fix still -5.6% > from your parent comparison point? Yes. > If we value that microbenchmark > at all (debatable), I'd say that's not good enough. I think that is a good improvement. > It does match with my own rough measurement, but I'd been hoping > for better when done in a more controlled environment; and I cannot > explain why "truncate prealloc blocks past i_size" creat-clo performance > would not be fully corrected by "avoid a little creat and stat slowdown" > (unless either patch adds subtle icache or dcache displacements). > > I'm not certain of how you performed the comparison. Was the > c435a390574d tree measured, then patch afa2db2fb6f1 applied on top > of that and measured, then patch 43819159da2b applied on top of that > and measured? Or were there other intervening changes, which could > easily add their own interference? c435a390574d is the direct parent of afa2db2fb6f1 in its original git. 43819159da2b is your patch applied on top of v4.3-rc7. The comparison of 43819159da2b with v4.3-rc7 is as follow: = compiler/cpufreq_governor/kconfig/rootfs/tbox_group/test/testcase/testtime: gcc-4.9/performance/x86_64-rhel/debian-x86_64-2015-02-07.cgz/lkp-wsx02/creat-clo/aim9/300s commit: 32b88194f71d6ae7768a29f87fbba454728273ee 43819159da2b77fedcf7562134d6003dccd6a068 32b88194f71d6ae7 43819159da2b77fedcf7562134 -- %stddev %change %stddev \ |\ 475224 ± 1% +11.9% 531968 ± 1% aim9.creat-clo.ops_per_sec 10469094 ±201% -52.3%4998529 ±130% latency_stats.avg.nfs_wait_on_request.nfs_updatepage.nfs_write_end.generic_perform_write.__generic_file_write_iter.generic_file_write_iter.nfs_file_write.__vfs_write.vfs_write.SyS_write.entry_SYSCALL_64_fastpath 18852332 ±223% -73.5%4998529 ±130% latency_stats.max.nfs_wait_on_request.nfs_updatepage.nfs_write_end.generic_perform_write.__generic_file_write_iter.generic_file_write_iter.nfs_file_write.__vfs_write.vfs_write.SyS_write.entry_SYSCALL_64_fastpath 21758590 ±199% -77.0%4998529 ±130% latency_stats.sum.nfs_wait_on_request.nfs_updatepage.nfs_write_end.generic_perform_write.__generic_file_write_iter.generic_file_write_iter.nfs_file_write.__vfs_write.vfs_write.SyS_write.entry_SYSCALL_64_fastpath 4817724 ± 0% +9.6%5280303 ± 1% proc-vmstat.numa_hit 4812582 ± 0% +9.7%5280287 ± 1% proc-vmstat.numa_local 8499767 ± 4% +14.2%9707953 ± 4% proc-vmstat.pgalloc_normal 8984075 ± 0% +10.4%9919044 ± 1% proc-vmstat.pgfree 9.22 ± 8% +27.4% 11.75 ± 9%
Re: [PATCH] tmpfs: avoid a little creat and stat slowdown
On Fri, 13 Nov 2015, Huang, Ying wrote: > > c435a390574d is the direct parent of afa2db2fb6f1 in its original git. > 43819159da2b is your patch applied on top of v4.3-rc7. The comparison > of 43819159da2b with v4.3-rc7 is as follow: ... > So you patch improved 11.9% from its base v4.3-rc7. I think other > difference are caused by other changes. Sorry for confusing. Thanks for getting back on this: that's rather what I was hoping to hear. Of course, no user will care which commit is responsible for a slowdown, and we may need to look further; but I couldn't make sense of it before, so this was a relief. Hugh -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] tmpfs: avoid a little creat and stat slowdown
On Wed, 4 Nov 2015, Huang, Ying wrote: > Hugh Dickins writes: > > > LKP reports that v4.2 commit afa2db2fb6f1 ("tmpfs: truncate prealloc > > blocks past i_size") causes a 14.5% slowdown in the AIM9 creat-clo > > benchmark. > > > > creat-clo does just what you'd expect from the name, and creat's O_TRUNC > > on 0-length file does indeed get into more overhead now shmem_setattr() > > tests "0 <= 0" instead of "0 < 0". > > > > I'm not sure how much we care, but I think it would not be too VW-like > > to add in a check for whether any pages (or swap) are allocated: if none > > are allocated, there's none to remove from the radix_tree. At first I > > thought that check would be good enough for the unmaps too, but no: we > > should not skip the unlikely case of unmapping pages beyond the new EOF, > > which were COWed from holes which have now been reclaimed, leaving none. > > > > This gives me an 8.5% speedup: on Haswell instead of LKP's Westmere, > > and running a debug config before and after: I hope those account for > > the lesser speedup. > > > > And probably someone has a benchmark where a thousand threads keep on > > stat'ing the same file repeatedly: forestall that report by adjusting > > v4.3 commit 44a30220bc0a ("shmem: recalculate file inode when fstat") > > not to take the spinlock in shmem_getattr() when there's no work to do. > > > > Reported-by: Ying Huang > > Signed-off-by: Hugh Dickins > > Hi, Hugh, > > Thanks a lot for your support! The test on LKP shows that this patch > restores a big part of the regression! In following list, > > c435a390574d012f8d30074135d8fcc6f480b484: is parent commit > afa2db2fb6f15f860069de94a1257db57589fe95: is the first bad commit has > performance regression. > 43819159da2b77fedcf7562134d6003dccd6a068: is the fixing patch Hi Ying, Thank you, for reporting, and for trying out the patch (which is now in Linus's tree as commit d0424c429f8e0555a337d71e0a13f2289c636ec9). But I'm disappointed by the result: do I understand correctly, that afa2db2fb6f1 made a -12.5% change, but the fix still -5.6% from your parent comparison point? If we value that microbenchmark at all (debatable), I'd say that's not good enough. It does match with my own rough measurement, but I'd been hoping for better when done in a more controlled environment; and I cannot explain why "truncate prealloc blocks past i_size" creat-clo performance would not be fully corrected by "avoid a little creat and stat slowdown" (unless either patch adds subtle icache or dcache displacements). I'm not certain of how you performed the comparison. Was the c435a390574d tree measured, then patch afa2db2fb6f1 applied on top of that and measured, then patch 43819159da2b applied on top of that and measured? Or were there other intervening changes, which could easily add their own interference? Hugh > > = > compiler/cpufreq_governor/kconfig/rootfs/tbox_group/test/testcase/testtime: > > gcc-4.9/performance/x86_64-rhel/debian-x86_64-2015-02-07.cgz/lkp-wsx02/creat-clo/aim9/300s > > commit: > c435a390574d012f8d30074135d8fcc6f480b484 > afa2db2fb6f15f860069de94a1257db57589fe95 > 43819159da2b77fedcf7562134d6003dccd6a068 > > c435a390574d012f afa2db2fb6f15f860069de94a1 43819159da2b77fedcf7562134 > -- -- > %stddev %change %stddev %change %stddev > \ |\ |\ > 563556 ± 1% -12.5% 493033 ± 5% -5.6% 531968 ± 1% > aim9.creat-clo.ops_per_sec > 11836 ± 7% +11.4% 13184 ± 7% +15.0% 13608 ± 5% > numa-meminfo.node1.SReclaimable > 10121526 ± 3% -12.1%8897097 ± 5% -4.1%9707953 ± 4% > proc-vmstat.pgalloc_normal > 9.34 ± 4% -11.4% 8.28 ± 3% -4.8% 8.88 ± 2% > time.user_time > 3480 ± 3% -2.5% 3395 ± 1% -28.5% 2488 ± 3% > vmstat.system.cs > 203275 ± 17% -6.8% 189453 ± 5% -34.4% 133352 ± 11% > cpuidle.C1-NHM.usage >8081280 ±129% -93.3% 538377 ± 97% +31.5% 10625496 ±106% > cpuidle.C1E-NHM.time > 3144 ± 58%+619.0% 22606 ± 56%+903.9% 31563 ± 0% > numa-vmstat.node0.numa_other > 2958 ± 7% +11.4% 3295 ± 7% +15.0% 3401 ± 5% > numa-vmstat.node1.nr_slab_reclaimable > 45074 ± 5% -43.4% 25494 ± 57% -68.7% 14105 ± 2% > numa-vmstat.node2.numa_other > 56140 ± 0% +0.0% 56158 ± 0% -94.4% 3120 ± 0% > slabinfo.Acpi-ParseExt.active_objs > 1002 ± 0% +0.0% 1002 ± 0% -92.0% 80.00 ± 0% > slabinfo.Acpi-ParseExt.active_slabs > 56140 ± 0% +0.0% 56158 ± 0% -94.4% 3120 ± 0% > slabinfo.Acpi-ParseExt.num_objs > 1002 ± 0% +0.0%
Re: [PATCH] tmpfs: avoid a little creat and stat slowdown
On Wed, 4 Nov 2015, Huang, Ying wrote: > Hugh Dickinswrites: > > > LKP reports that v4.2 commit afa2db2fb6f1 ("tmpfs: truncate prealloc > > blocks past i_size") causes a 14.5% slowdown in the AIM9 creat-clo > > benchmark. > > > > creat-clo does just what you'd expect from the name, and creat's O_TRUNC > > on 0-length file does indeed get into more overhead now shmem_setattr() > > tests "0 <= 0" instead of "0 < 0". > > > > I'm not sure how much we care, but I think it would not be too VW-like > > to add in a check for whether any pages (or swap) are allocated: if none > > are allocated, there's none to remove from the radix_tree. At first I > > thought that check would be good enough for the unmaps too, but no: we > > should not skip the unlikely case of unmapping pages beyond the new EOF, > > which were COWed from holes which have now been reclaimed, leaving none. > > > > This gives me an 8.5% speedup: on Haswell instead of LKP's Westmere, > > and running a debug config before and after: I hope those account for > > the lesser speedup. > > > > And probably someone has a benchmark where a thousand threads keep on > > stat'ing the same file repeatedly: forestall that report by adjusting > > v4.3 commit 44a30220bc0a ("shmem: recalculate file inode when fstat") > > not to take the spinlock in shmem_getattr() when there's no work to do. > > > > Reported-by: Ying Huang > > Signed-off-by: Hugh Dickins > > Hi, Hugh, > > Thanks a lot for your support! The test on LKP shows that this patch > restores a big part of the regression! In following list, > > c435a390574d012f8d30074135d8fcc6f480b484: is parent commit > afa2db2fb6f15f860069de94a1257db57589fe95: is the first bad commit has > performance regression. > 43819159da2b77fedcf7562134d6003dccd6a068: is the fixing patch Hi Ying, Thank you, for reporting, and for trying out the patch (which is now in Linus's tree as commit d0424c429f8e0555a337d71e0a13f2289c636ec9). But I'm disappointed by the result: do I understand correctly, that afa2db2fb6f1 made a -12.5% change, but the fix still -5.6% from your parent comparison point? If we value that microbenchmark at all (debatable), I'd say that's not good enough. It does match with my own rough measurement, but I'd been hoping for better when done in a more controlled environment; and I cannot explain why "truncate prealloc blocks past i_size" creat-clo performance would not be fully corrected by "avoid a little creat and stat slowdown" (unless either patch adds subtle icache or dcache displacements). I'm not certain of how you performed the comparison. Was the c435a390574d tree measured, then patch afa2db2fb6f1 applied on top of that and measured, then patch 43819159da2b applied on top of that and measured? Or were there other intervening changes, which could easily add their own interference? Hugh > > = > compiler/cpufreq_governor/kconfig/rootfs/tbox_group/test/testcase/testtime: > > gcc-4.9/performance/x86_64-rhel/debian-x86_64-2015-02-07.cgz/lkp-wsx02/creat-clo/aim9/300s > > commit: > c435a390574d012f8d30074135d8fcc6f480b484 > afa2db2fb6f15f860069de94a1257db57589fe95 > 43819159da2b77fedcf7562134d6003dccd6a068 > > c435a390574d012f afa2db2fb6f15f860069de94a1 43819159da2b77fedcf7562134 > -- -- > %stddev %change %stddev %change %stddev > \ |\ |\ > 563556 ± 1% -12.5% 493033 ± 5% -5.6% 531968 ± 1% > aim9.creat-clo.ops_per_sec > 11836 ± 7% +11.4% 13184 ± 7% +15.0% 13608 ± 5% > numa-meminfo.node1.SReclaimable > 10121526 ± 3% -12.1%8897097 ± 5% -4.1%9707953 ± 4% > proc-vmstat.pgalloc_normal > 9.34 ± 4% -11.4% 8.28 ± 3% -4.8% 8.88 ± 2% > time.user_time > 3480 ± 3% -2.5% 3395 ± 1% -28.5% 2488 ± 3% > vmstat.system.cs > 203275 ± 17% -6.8% 189453 ± 5% -34.4% 133352 ± 11% > cpuidle.C1-NHM.usage >8081280 ±129% -93.3% 538377 ± 97% +31.5% 10625496 ±106% > cpuidle.C1E-NHM.time > 3144 ± 58%+619.0% 22606 ± 56%+903.9% 31563 ± 0% > numa-vmstat.node0.numa_other > 2958 ± 7% +11.4% 3295 ± 7% +15.0% 3401 ± 5% > numa-vmstat.node1.nr_slab_reclaimable > 45074 ± 5% -43.4% 25494 ± 57% -68.7% 14105 ± 2% > numa-vmstat.node2.numa_other > 56140 ± 0% +0.0% 56158 ± 0% -94.4% 3120 ± 0% > slabinfo.Acpi-ParseExt.active_objs > 1002 ± 0% +0.0% 1002 ± 0% -92.0% 80.00 ± 0% > slabinfo.Acpi-ParseExt.active_slabs > 56140 ± 0% +0.0% 56158 ± 0% -94.4% 3120 ± 0% >
Re: [PATCH] tmpfs: avoid a little creat and stat slowdown
Hugh Dickins writes: > LKP reports that v4.2 commit afa2db2fb6f1 ("tmpfs: truncate prealloc > blocks past i_size") causes a 14.5% slowdown in the AIM9 creat-clo > benchmark. > > creat-clo does just what you'd expect from the name, and creat's O_TRUNC > on 0-length file does indeed get into more overhead now shmem_setattr() > tests "0 <= 0" instead of "0 < 0". > > I'm not sure how much we care, but I think it would not be too VW-like > to add in a check for whether any pages (or swap) are allocated: if none > are allocated, there's none to remove from the radix_tree. At first I > thought that check would be good enough for the unmaps too, but no: we > should not skip the unlikely case of unmapping pages beyond the new EOF, > which were COWed from holes which have now been reclaimed, leaving none. > > This gives me an 8.5% speedup: on Haswell instead of LKP's Westmere, > and running a debug config before and after: I hope those account for > the lesser speedup. > > And probably someone has a benchmark where a thousand threads keep on > stat'ing the same file repeatedly: forestall that report by adjusting > v4.3 commit 44a30220bc0a ("shmem: recalculate file inode when fstat") > not to take the spinlock in shmem_getattr() when there's no work to do. > > Reported-by: Ying Huang > Signed-off-by: Hugh Dickins Hi, Hugh, Thanks a lot for your support! The test on LKP shows that this patch restores a big part of the regression! In following list, c435a390574d012f8d30074135d8fcc6f480b484: is parent commit afa2db2fb6f15f860069de94a1257db57589fe95: is the first bad commit has performance regression. 43819159da2b77fedcf7562134d6003dccd6a068: is the fixing patch = compiler/cpufreq_governor/kconfig/rootfs/tbox_group/test/testcase/testtime: gcc-4.9/performance/x86_64-rhel/debian-x86_64-2015-02-07.cgz/lkp-wsx02/creat-clo/aim9/300s commit: c435a390574d012f8d30074135d8fcc6f480b484 afa2db2fb6f15f860069de94a1257db57589fe95 43819159da2b77fedcf7562134d6003dccd6a068 c435a390574d012f afa2db2fb6f15f860069de94a1 43819159da2b77fedcf7562134 -- -- %stddev %change %stddev %change %stddev \ |\ |\ 563556 ± 1% -12.5% 493033 ± 5% -5.6% 531968 ± 1% aim9.creat-clo.ops_per_sec 11836 ± 7% +11.4% 13184 ± 7% +15.0% 13608 ± 5% numa-meminfo.node1.SReclaimable 10121526 ± 3% -12.1%8897097 ± 5% -4.1%9707953 ± 4% proc-vmstat.pgalloc_normal 9.34 ± 4% -11.4% 8.28 ± 3% -4.8% 8.88 ± 2% time.user_time 3480 ± 3% -2.5% 3395 ± 1% -28.5% 2488 ± 3% vmstat.system.cs 203275 ± 17% -6.8% 189453 ± 5% -34.4% 133352 ± 11% cpuidle.C1-NHM.usage 8081280 ±129% -93.3% 538377 ± 97% +31.5% 10625496 ±106% cpuidle.C1E-NHM.time 3144 ± 58%+619.0% 22606 ± 56%+903.9% 31563 ± 0% numa-vmstat.node0.numa_other 2958 ± 7% +11.4% 3295 ± 7% +15.0% 3401 ± 5% numa-vmstat.node1.nr_slab_reclaimable 45074 ± 5% -43.4% 25494 ± 57% -68.7% 14105 ± 2% numa-vmstat.node2.numa_other 56140 ± 0% +0.0% 56158 ± 0% -94.4% 3120 ± 0% slabinfo.Acpi-ParseExt.active_objs 1002 ± 0% +0.0% 1002 ± 0% -92.0% 80.00 ± 0% slabinfo.Acpi-ParseExt.active_slabs 56140 ± 0% +0.0% 56158 ± 0% -94.4% 3120 ± 0% slabinfo.Acpi-ParseExt.num_objs 1002 ± 0% +0.0% 1002 ± 0% -92.0% 80.00 ± 0% slabinfo.Acpi-ParseExt.num_slabs 1079 ± 5% -10.8% 962.00 ± 10%-100.0% 0.00 ± -1% slabinfo.blkdev_ioc.active_objs 1079 ± 5% -10.8% 962.00 ± 10%-100.0% 0.00 ± -1% slabinfo.blkdev_ioc.num_objs 110.67 ± 39% +74.4% 193.00 ± 46%+317.5% 462.00 ± 8% slabinfo.blkdev_queue.active_objs 189.33 ± 23% +43.7% 272.00 ± 33%+151.4% 476.00 ± 10% slabinfo.blkdev_queue.num_objs 1129 ± 10% -1.9% 1107 ± 7% +20.8% 1364 ± 6% slabinfo.blkdev_requests.active_objs 1129 ± 10% -1.9% 1107 ± 7% +20.8% 1364 ± 6% slabinfo.blkdev_requests.num_objs 1058 ± 3% -10.3% 949.00 ± 9%-100.0% 0.00 ± -1% slabinfo.file_lock_ctx.active_objs 1058 ± 3% -10.3% 949.00 ± 9%-100.0% 0.00 ± -1% slabinfo.file_lock_ctx.num_objs 4060 ± 1% -2.1% 3973 ± 1% -10.5% 3632 ± 1% slabinfo.files_cache.active_objs 4060 ± 1% -2.1% 3973 ± 1% -10.5% 3632 ± 1% slabinfo.files_cache.num_objs 10001 ± 0% -0.3% 9973 ± 0% -61.1% 3888 ± 0%
Re: [PATCH] tmpfs: avoid a little creat and stat slowdown
Hugh Dickinswrites: > LKP reports that v4.2 commit afa2db2fb6f1 ("tmpfs: truncate prealloc > blocks past i_size") causes a 14.5% slowdown in the AIM9 creat-clo > benchmark. > > creat-clo does just what you'd expect from the name, and creat's O_TRUNC > on 0-length file does indeed get into more overhead now shmem_setattr() > tests "0 <= 0" instead of "0 < 0". > > I'm not sure how much we care, but I think it would not be too VW-like > to add in a check for whether any pages (or swap) are allocated: if none > are allocated, there's none to remove from the radix_tree. At first I > thought that check would be good enough for the unmaps too, but no: we > should not skip the unlikely case of unmapping pages beyond the new EOF, > which were COWed from holes which have now been reclaimed, leaving none. > > This gives me an 8.5% speedup: on Haswell instead of LKP's Westmere, > and running a debug config before and after: I hope those account for > the lesser speedup. > > And probably someone has a benchmark where a thousand threads keep on > stat'ing the same file repeatedly: forestall that report by adjusting > v4.3 commit 44a30220bc0a ("shmem: recalculate file inode when fstat") > not to take the spinlock in shmem_getattr() when there's no work to do. > > Reported-by: Ying Huang > Signed-off-by: Hugh Dickins Hi, Hugh, Thanks a lot for your support! The test on LKP shows that this patch restores a big part of the regression! In following list, c435a390574d012f8d30074135d8fcc6f480b484: is parent commit afa2db2fb6f15f860069de94a1257db57589fe95: is the first bad commit has performance regression. 43819159da2b77fedcf7562134d6003dccd6a068: is the fixing patch = compiler/cpufreq_governor/kconfig/rootfs/tbox_group/test/testcase/testtime: gcc-4.9/performance/x86_64-rhel/debian-x86_64-2015-02-07.cgz/lkp-wsx02/creat-clo/aim9/300s commit: c435a390574d012f8d30074135d8fcc6f480b484 afa2db2fb6f15f860069de94a1257db57589fe95 43819159da2b77fedcf7562134d6003dccd6a068 c435a390574d012f afa2db2fb6f15f860069de94a1 43819159da2b77fedcf7562134 -- -- %stddev %change %stddev %change %stddev \ |\ |\ 563556 ± 1% -12.5% 493033 ± 5% -5.6% 531968 ± 1% aim9.creat-clo.ops_per_sec 11836 ± 7% +11.4% 13184 ± 7% +15.0% 13608 ± 5% numa-meminfo.node1.SReclaimable 10121526 ± 3% -12.1%8897097 ± 5% -4.1%9707953 ± 4% proc-vmstat.pgalloc_normal 9.34 ± 4% -11.4% 8.28 ± 3% -4.8% 8.88 ± 2% time.user_time 3480 ± 3% -2.5% 3395 ± 1% -28.5% 2488 ± 3% vmstat.system.cs 203275 ± 17% -6.8% 189453 ± 5% -34.4% 133352 ± 11% cpuidle.C1-NHM.usage 8081280 ±129% -93.3% 538377 ± 97% +31.5% 10625496 ±106% cpuidle.C1E-NHM.time 3144 ± 58%+619.0% 22606 ± 56%+903.9% 31563 ± 0% numa-vmstat.node0.numa_other 2958 ± 7% +11.4% 3295 ± 7% +15.0% 3401 ± 5% numa-vmstat.node1.nr_slab_reclaimable 45074 ± 5% -43.4% 25494 ± 57% -68.7% 14105 ± 2% numa-vmstat.node2.numa_other 56140 ± 0% +0.0% 56158 ± 0% -94.4% 3120 ± 0% slabinfo.Acpi-ParseExt.active_objs 1002 ± 0% +0.0% 1002 ± 0% -92.0% 80.00 ± 0% slabinfo.Acpi-ParseExt.active_slabs 56140 ± 0% +0.0% 56158 ± 0% -94.4% 3120 ± 0% slabinfo.Acpi-ParseExt.num_objs 1002 ± 0% +0.0% 1002 ± 0% -92.0% 80.00 ± 0% slabinfo.Acpi-ParseExt.num_slabs 1079 ± 5% -10.8% 962.00 ± 10%-100.0% 0.00 ± -1% slabinfo.blkdev_ioc.active_objs 1079 ± 5% -10.8% 962.00 ± 10%-100.0% 0.00 ± -1% slabinfo.blkdev_ioc.num_objs 110.67 ± 39% +74.4% 193.00 ± 46%+317.5% 462.00 ± 8% slabinfo.blkdev_queue.active_objs 189.33 ± 23% +43.7% 272.00 ± 33%+151.4% 476.00 ± 10% slabinfo.blkdev_queue.num_objs 1129 ± 10% -1.9% 1107 ± 7% +20.8% 1364 ± 6% slabinfo.blkdev_requests.active_objs 1129 ± 10% -1.9% 1107 ± 7% +20.8% 1364 ± 6% slabinfo.blkdev_requests.num_objs 1058 ± 3% -10.3% 949.00 ± 9%-100.0% 0.00 ± -1% slabinfo.file_lock_ctx.active_objs 1058 ± 3% -10.3% 949.00 ± 9%-100.0% 0.00 ± -1% slabinfo.file_lock_ctx.num_objs 4060 ± 1% -2.1% 3973 ± 1% -10.5% 3632 ± 1% slabinfo.files_cache.active_objs 4060 ± 1% -2.1% 3973 ± 1% -10.5% 3632 ± 1% slabinfo.files_cache.num_objs 10001 ± 0% -0.3% 9973
[PATCH] tmpfs: avoid a little creat and stat slowdown
LKP reports that v4.2 commit afa2db2fb6f1 ("tmpfs: truncate prealloc blocks past i_size") causes a 14.5% slowdown in the AIM9 creat-clo benchmark. creat-clo does just what you'd expect from the name, and creat's O_TRUNC on 0-length file does indeed get into more overhead now shmem_setattr() tests "0 <= 0" instead of "0 < 0". I'm not sure how much we care, but I think it would not be too VW-like to add in a check for whether any pages (or swap) are allocated: if none are allocated, there's none to remove from the radix_tree. At first I thought that check would be good enough for the unmaps too, but no: we should not skip the unlikely case of unmapping pages beyond the new EOF, which were COWed from holes which have now been reclaimed, leaving none. This gives me an 8.5% speedup: on Haswell instead of LKP's Westmere, and running a debug config before and after: I hope those account for the lesser speedup. And probably someone has a benchmark where a thousand threads keep on stat'ing the same file repeatedly: forestall that report by adjusting v4.3 commit 44a30220bc0a ("shmem: recalculate file inode when fstat") not to take the spinlock in shmem_getattr() when there's no work to do. Reported-by: Ying Huang Signed-off-by: Hugh Dickins --- mm/shmem.c | 22 ++ 1 file changed, 14 insertions(+), 8 deletions(-) --- 4.3-rc7/mm/shmem.c 2015-09-12 18:30:20.857039763 -0700 +++ linux/mm/shmem.c2015-10-25 11:49:19.931973850 -0700 @@ -548,12 +548,12 @@ static int shmem_getattr(struct vfsmount struct inode *inode = dentry->d_inode; struct shmem_inode_info *info = SHMEM_I(inode); - spin_lock(>lock); - shmem_recalc_inode(inode); - spin_unlock(>lock); - + if (info->alloced - info->swapped != inode->i_mapping->nrpages) { + spin_lock(>lock); + shmem_recalc_inode(inode); + spin_unlock(>lock); + } generic_fillattr(inode, stat); - return 0; } @@ -586,10 +586,16 @@ static int shmem_setattr(struct dentry * } if (newsize <= oldsize) { loff_t holebegin = round_up(newsize, PAGE_SIZE); - unmap_mapping_range(inode->i_mapping, holebegin, 0, 1); - shmem_truncate_range(inode, newsize, (loff_t)-1); + if (oldsize > holebegin) + unmap_mapping_range(inode->i_mapping, + holebegin, 0, 1); + if (info->alloced) + shmem_truncate_range(inode, + newsize, (loff_t)-1); /* unmap again to remove racily COWed private pages */ - unmap_mapping_range(inode->i_mapping, holebegin, 0, 1); + if (oldsize > holebegin) + unmap_mapping_range(inode->i_mapping, + holebegin, 0, 1); } } -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[PATCH] tmpfs: avoid a little creat and stat slowdown
LKP reports that v4.2 commit afa2db2fb6f1 ("tmpfs: truncate prealloc blocks past i_size") causes a 14.5% slowdown in the AIM9 creat-clo benchmark. creat-clo does just what you'd expect from the name, and creat's O_TRUNC on 0-length file does indeed get into more overhead now shmem_setattr() tests "0 <= 0" instead of "0 < 0". I'm not sure how much we care, but I think it would not be too VW-like to add in a check for whether any pages (or swap) are allocated: if none are allocated, there's none to remove from the radix_tree. At first I thought that check would be good enough for the unmaps too, but no: we should not skip the unlikely case of unmapping pages beyond the new EOF, which were COWed from holes which have now been reclaimed, leaving none. This gives me an 8.5% speedup: on Haswell instead of LKP's Westmere, and running a debug config before and after: I hope those account for the lesser speedup. And probably someone has a benchmark where a thousand threads keep on stat'ing the same file repeatedly: forestall that report by adjusting v4.3 commit 44a30220bc0a ("shmem: recalculate file inode when fstat") not to take the spinlock in shmem_getattr() when there's no work to do. Reported-by: Ying HuangSigned-off-by: Hugh Dickins --- mm/shmem.c | 22 ++ 1 file changed, 14 insertions(+), 8 deletions(-) --- 4.3-rc7/mm/shmem.c 2015-09-12 18:30:20.857039763 -0700 +++ linux/mm/shmem.c2015-10-25 11:49:19.931973850 -0700 @@ -548,12 +548,12 @@ static int shmem_getattr(struct vfsmount struct inode *inode = dentry->d_inode; struct shmem_inode_info *info = SHMEM_I(inode); - spin_lock(>lock); - shmem_recalc_inode(inode); - spin_unlock(>lock); - + if (info->alloced - info->swapped != inode->i_mapping->nrpages) { + spin_lock(>lock); + shmem_recalc_inode(inode); + spin_unlock(>lock); + } generic_fillattr(inode, stat); - return 0; } @@ -586,10 +586,16 @@ static int shmem_setattr(struct dentry * } if (newsize <= oldsize) { loff_t holebegin = round_up(newsize, PAGE_SIZE); - unmap_mapping_range(inode->i_mapping, holebegin, 0, 1); - shmem_truncate_range(inode, newsize, (loff_t)-1); + if (oldsize > holebegin) + unmap_mapping_range(inode->i_mapping, + holebegin, 0, 1); + if (info->alloced) + shmem_truncate_range(inode, + newsize, (loff_t)-1); /* unmap again to remove racily COWed private pages */ - unmap_mapping_range(inode->i_mapping, holebegin, 0, 1); + if (oldsize > holebegin) + unmap_mapping_range(inode->i_mapping, + holebegin, 0, 1); } } -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/