/openbmc/linux/arch/x86/include/asm/ |
H A D | tlb.h | 611ae8e3 Wed Jun 27 20:02:22 CDT 2012 Alex Shi <alex.shi@intel.com> x86/tlb: enable tlb flush range support for x86
Not every tlb_flush execution moment is really need to evacuate all TLB entries, like in munmap, just few 'invlpg' is better for whole process performance, since it leaves most of TLB entries for later accessing.
This patch also rewrite flush_tlb_range for 2 purposes: 1, split it out to get flush_blt_mm_range function. 2, clean up to reduce line breaking, thanks for Borislav's input.
My micro benchmark 'mummap' http://lkml.org/lkml/2012/5/17/59 show that the random memory access on other CPU has 0~50% speed up on a 2P * 4cores * HT NHM EP while do 'munmap'.
Thanks Yongjie's testing on this patch: ------------- I used Linux 3.4-RC6 w/ and w/o his patches as Xen dom0 and guest kernel. After running two benchmarks in Xen HVM guest, I found his patches brought about 1%~3% performance gain in 'kernel build' and 'netperf' testing, though the performance gain was not very stable in 'kernel build' testing.
Some detailed testing results are below.
Testing Environment: Hardware: Romley-EP platform Xen version: latest upstream Linux kernel: 3.4-RC6 Guest vCPU number: 8 NIC: Intel 82599 (10GB bandwidth)
In 'kernel build' testing in guest: Command line | performance gain make -j 4 | 3.81% make -j 8 | 0.37% make -j 16 | -0.52%
In 'netperf' testing, we tested TCP_STREAM with default socket size 16384 byte as large packet and 64 byte as small packet. I used several clients to add networking pressure, then 'netperf' server automatically generated several threads to response them. I also used large-size packet and small-size packet in the testing. Packet size | Thread number | performance gain 16384 bytes | 4 | 0.02% 16384 bytes | 8 | 2.21% 16384 bytes | 16 | 2.04% 64 bytes | 4 | 1.07% 64 bytes | 8 | 3.31% 64 bytes | 16 | 0.71%
Signed-off-by: Alex Shi <alex.shi@intel.com> Link: http://lkml.kernel.org/r/1340845344-27557-8-git-send-email-alex.shi@intel.com Tested-by: Ren, Yongjie <yongjie.ren@intel.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com> 611ae8e3 Wed Jun 27 20:02:22 CDT 2012 Alex Shi <alex.shi@intel.com> x86/tlb: enable tlb flush range support for x86 Not every tlb_flush execution moment is really need to evacuate all TLB entries, like in munmap, just few 'invlpg' is better for whole process performance, since it leaves most of TLB entries for later accessing. This patch also rewrite flush_tlb_range for 2 purposes: 1, split it out to get flush_blt_mm_range function. 2, clean up to reduce line breaking, thanks for Borislav's input. My micro benchmark 'mummap' http://lkml.org/lkml/2012/5/17/59 show that the random memory access on other CPU has 0~50% speed up on a 2P * 4cores * HT NHM EP while do 'munmap'. Thanks Yongjie's testing on this patch: ------------- I used Linux 3.4-RC6 w/ and w/o his patches as Xen dom0 and guest kernel. After running two benchmarks in Xen HVM guest, I found his patches brought about 1%~3% performance gain in 'kernel build' and 'netperf' testing, though the performance gain was not very stable in 'kernel build' testing. Some detailed testing results are below. Testing Environment: Hardware: Romley-EP platform Xen version: latest upstream Linux kernel: 3.4-RC6 Guest vCPU number: 8 NIC: Intel 82599 (10GB bandwidth) In 'kernel build' testing in guest: Command line | performance gain make -j 4 | 3.81% make -j 8 | 0.37% make -j 16 | -0.52% In 'netperf' testing, we tested TCP_STREAM with default socket size 16384 byte as large packet and 64 byte as small packet. I used several clients to add networking pressure, then 'netperf' server automatically generated several threads to response them. I also used large-size packet and small-size packet in the testing. Packet size | Thread number | performance gain 16384 bytes | 4 | 0.02% 16384 bytes | 8 | 2.21% 16384 bytes | 16 | 2.04% 64 bytes | 4 | 1.07% 64 bytes | 8 | 3.31% 64 bytes | 16 | 0.71% Signed-off-by: Alex Shi <alex.shi@intel.com> Link: http://lkml.kernel.org/r/1340845344-27557-8-git-send-email-alex.shi@intel.com Tested-by: Ren, Yongjie <yongjie.ren@intel.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
|
H A D | tlbflush.h | 611ae8e3 Wed Jun 27 20:02:22 CDT 2012 Alex Shi <alex.shi@intel.com> x86/tlb: enable tlb flush range support for x86
Not every tlb_flush execution moment is really need to evacuate all TLB entries, like in munmap, just few 'invlpg' is better for whole process performance, since it leaves most of TLB entries for later accessing.
This patch also rewrite flush_tlb_range for 2 purposes: 1, split it out to get flush_blt_mm_range function. 2, clean up to reduce line breaking, thanks for Borislav's input.
My micro benchmark 'mummap' http://lkml.org/lkml/2012/5/17/59 show that the random memory access on other CPU has 0~50% speed up on a 2P * 4cores * HT NHM EP while do 'munmap'.
Thanks Yongjie's testing on this patch: ------------- I used Linux 3.4-RC6 w/ and w/o his patches as Xen dom0 and guest kernel. After running two benchmarks in Xen HVM guest, I found his patches brought about 1%~3% performance gain in 'kernel build' and 'netperf' testing, though the performance gain was not very stable in 'kernel build' testing.
Some detailed testing results are below.
Testing Environment: Hardware: Romley-EP platform Xen version: latest upstream Linux kernel: 3.4-RC6 Guest vCPU number: 8 NIC: Intel 82599 (10GB bandwidth)
In 'kernel build' testing in guest: Command line | performance gain make -j 4 | 3.81% make -j 8 | 0.37% make -j 16 | -0.52%
In 'netperf' testing, we tested TCP_STREAM with default socket size 16384 byte as large packet and 64 byte as small packet. I used several clients to add networking pressure, then 'netperf' server automatically generated several threads to response them. I also used large-size packet and small-size packet in the testing. Packet size | Thread number | performance gain 16384 bytes | 4 | 0.02% 16384 bytes | 8 | 2.21% 16384 bytes | 16 | 2.04% 64 bytes | 4 | 1.07% 64 bytes | 8 | 3.31% 64 bytes | 16 | 0.71%
Signed-off-by: Alex Shi <alex.shi@intel.com> Link: http://lkml.kernel.org/r/1340845344-27557-8-git-send-email-alex.shi@intel.com Tested-by: Ren, Yongjie <yongjie.ren@intel.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com> 611ae8e3 Wed Jun 27 20:02:22 CDT 2012 Alex Shi <alex.shi@intel.com> x86/tlb: enable tlb flush range support for x86 Not every tlb_flush execution moment is really need to evacuate all TLB entries, like in munmap, just few 'invlpg' is better for whole process performance, since it leaves most of TLB entries for later accessing. This patch also rewrite flush_tlb_range for 2 purposes: 1, split it out to get flush_blt_mm_range function. 2, clean up to reduce line breaking, thanks for Borislav's input. My micro benchmark 'mummap' http://lkml.org/lkml/2012/5/17/59 show that the random memory access on other CPU has 0~50% speed up on a 2P * 4cores * HT NHM EP while do 'munmap'. Thanks Yongjie's testing on this patch: ------------- I used Linux 3.4-RC6 w/ and w/o his patches as Xen dom0 and guest kernel. After running two benchmarks in Xen HVM guest, I found his patches brought about 1%~3% performance gain in 'kernel build' and 'netperf' testing, though the performance gain was not very stable in 'kernel build' testing. Some detailed testing results are below. Testing Environment: Hardware: Romley-EP platform Xen version: latest upstream Linux kernel: 3.4-RC6 Guest vCPU number: 8 NIC: Intel 82599 (10GB bandwidth) In 'kernel build' testing in guest: Command line | performance gain make -j 4 | 3.81% make -j 8 | 0.37% make -j 16 | -0.52% In 'netperf' testing, we tested TCP_STREAM with default socket size 16384 byte as large packet and 64 byte as small packet. I used several clients to add networking pressure, then 'netperf' server automatically generated several threads to response them. I also used large-size packet and small-size packet in the testing. Packet size | Thread number | performance gain 16384 bytes | 4 | 0.02% 16384 bytes | 8 | 2.21% 16384 bytes | 16 | 2.04% 64 bytes | 4 | 1.07% 64 bytes | 8 | 3.31% 64 bytes | 16 | 0.71% Signed-off-by: Alex Shi <alex.shi@intel.com> Link: http://lkml.kernel.org/r/1340845344-27557-8-git-send-email-alex.shi@intel.com Tested-by: Ren, Yongjie <yongjie.ren@intel.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
|
/openbmc/linux/arch/x86/kernel/cpu/ |
H A D | intel.c | b9a3b4c9 Tue Jan 21 16:33:22 CST 2014 Mel Gorman <mgorman@suse.de> mm, x86: Revisit tlb_flushall_shift tuning for page flushes except on IvyBridge
There was a large ebizzy performance regression that was bisected to commit 611ae8e3 (x86/tlb: enable tlb flush range support for x86). The problem was related to the tlb_flushall_shift tuning for IvyBridge which was altered. The problem is that it is not clear if the tuning values for each CPU family is correct as the methodology used to tune the values is unclear.
This patch uses a conservative tlb_flushall_shift value for all CPU families except IvyBridge so the decision can be revisited if any regression is found as a result of this change. IvyBridge is an exception as testing with one methodology determined that the value of 2 is acceptable. Details are in the changelog for the patch "x86: mm: Change tlb_flushall_shift for IvyBridge".
One important aspect of this to watch out for is Xen. The original commit log mentioned large performance gains on Xen. It's possible Xen is more sensitive to this value if it flushes small ranges of pages more frequently than workloads on bare metal typically do.
Signed-off-by: Mel Gorman <mgorman@suse.de> Tested-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Alex Shi <alex.shi@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/n/tip-dyzMww3fqugnhbhgo6Gxmtkw@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org> f98b7a77 Tue Jan 21 16:33:21 CST 2014 Mel Gorman <mgorman@suse.de> x86: mm: change tlb_flushall_shift for IvyBridge
There was a large performance regression that was bisected to commit 611ae8e3 ("x86/tlb: enable tlb flush range support for x86"). This patch simply changes the default balance point between a local and global flush for IvyBridge.
In the interest of allowing the tests to be reproduced, this patch was tested using mmtests 0.15 with the following configurations
configs/config-global-dhp__tlbflush-performance configs/config-global-dhp__scheduler-performance configs/config-global-dhp__network-performance
Results are from two machines
Ivybridge 4 threads: Intel(R) Core(TM) i3-3240 CPU @ 3.40GHz Ivybridge 8 threads: Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz
Page fault microbenchmark showed nothing interesting.
Ebizzy was configured to run multiple iterations and threads. Thread counts ranged from 1 to NR_CPUS*2. For each thread count, it ran 100 iterations and each iteration lasted 10 seconds.
Ivybridge 4 threads 3.13.0-rc7 3.13.0-rc7 vanilla altshift-v3 Mean 1 6395.44 ( 0.00%) 6789.09 ( 6.16%) Mean 2 7012.85 ( 0.00%) 8052.16 ( 14.82%) Mean 3 6403.04 ( 0.00%) 6973.74 ( 8.91%) Mean 4 6135.32 ( 0.00%) 6582.33 ( 7.29%) Mean 5 6095.69 ( 0.00%) 6526.68 ( 7.07%) Mean 6 6114.33 ( 0.00%) 6416.64 ( 4.94%) Mean 7 6085.10 ( 0.00%) 6448.51 ( 5.97%) Mean 8 6120.62 ( 0.00%) 6462.97 ( 5.59%)
Ivybridge 8 threads 3.13.0-rc7 3.13.0-rc7 vanilla altshift-v3 Mean 1 7336.65 ( 0.00%) 7787.02 ( 6.14%) Mean 2 8218.41 ( 0.00%) 9484.13 ( 15.40%) Mean 3 7973.62 ( 0.00%) 8922.01 ( 11.89%) Mean 4 7798.33 ( 0.00%) 8567.03 ( 9.86%) Mean 5 7158.72 ( 0.00%) 8214.23 ( 14.74%) Mean 6 6852.27 ( 0.00%) 7952.45 ( 16.06%) Mean 7 6774.65 ( 0.00%) 7536.35 ( 11.24%) Mean 8 6510.50 ( 0.00%) 6894.05 ( 5.89%) Mean 12 6182.90 ( 0.00%) 6661.29 ( 7.74%) Mean 16 6100.09 ( 0.00%) 6608.69 ( 8.34%)
Ebizzy hits the worst case scenario for TLB range flushing every time and it shows for these Ivybridge CPUs at least that the default choice is a poor on. The patch addresses the problem.
Next was a tlbflush microbenchmark written by Alex Shi at http://marc.info/?l=linux-kernel&m=133727348217113 . It measures access costs while the TLB is being flushed. The expectation is that if there are always full TLB flushes that the benchmark would suffer and it benefits from range flushing
There are 320 iterations of the test per thread count. The number of entries is randomly selected with a min of 1 and max of 512. To ensure a reasonably even spread of entries, the full range is broken up into 8 sections and a random number selected within that section.
iteration 1, random number between 0-64 iteration 2, random number between 64-128 etc
This is still a very weak methodology. When you do not know what are typical ranges, random is a reasonable choice but it can be easily argued that the opimisation was for smaller ranges and an even spread is not representative of any workload that matters. To improve this, we'd need to know the probability distribution of TLB flush range sizes for a set of workloads that are considered "common", build a synthetic trace and feed that into this benchmark. Even that is not perfect because it would not account for the time between flushes but there are limits of what can be reasonably done and still be doing something useful. If a representative synthetic trace is provided then this benchmark could be revisited and the shift values retuned.
Ivybridge 4 threads 3.13.0-rc7 3.13.0-rc7 vanilla altshift-v3 Mean 1 10.50 ( 0.00%) 10.50 ( 0.03%) Mean 2 17.59 ( 0.00%) 17.18 ( 2.34%) Mean 3 22.98 ( 0.00%) 21.74 ( 5.41%) Mean 5 47.13 ( 0.00%) 46.23 ( 1.92%) Mean 8 43.30 ( 0.00%) 42.56 ( 1.72%)
Ivybridge 8 threads 3.13.0-rc7 3.13.0-rc7 vanilla altshift-v3 Mean 1 9.45 ( 0.00%) 9.36 ( 0.93%) Mean 2 9.37 ( 0.00%) 9.70 ( -3.54%) Mean 3 9.36 ( 0.00%) 9.29 ( 0.70%) Mean 5 14.49 ( 0.00%) 15.04 ( -3.75%) Mean 8 41.08 ( 0.00%) 38.73 ( 5.71%) Mean 13 32.04 ( 0.00%) 31.24 ( 2.49%) Mean 16 40.05 ( 0.00%) 39.04 ( 2.51%)
For both CPUs, average access time is reduced which is good as this is the benchmark that was used to tune the shift values in the first place albeit it is now known *how* the benchmark was used.
The scheduler benchmarks were somewhat inconclusive. They showed gains and losses and makes me reconsider how stable those benchmarks really are or if something else might be interfering with the test results recently.
Network benchmarks were inconclusive. Almost all results were flat except for netperf-udp tests on the 4 thread machine. These results were unstable and showed large variations between reboots. It is unknown if this is a recent problems but I've noticed before that netperf-udp results tend to vary.
Based on these results, changing the default for Ivybridge seems like a logical choice.
Signed-off-by: Mel Gorman <mgorman@suse.de> Tested-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Alex Shi <alex.shi@linaro.org> Reviewed-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/n/tip-cqnadffh1tiqrshthRj3Esge@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org> b9a3b4c9 Tue Jan 21 16:33:22 CST 2014 Mel Gorman <mgorman@suse.de> mm, x86: Revisit tlb_flushall_shift tuning for page flushes except on IvyBridge There was a large ebizzy performance regression that was bisected to commit 611ae8e3 (x86/tlb: enable tlb flush range support for x86). The problem was related to the tlb_flushall_shift tuning for IvyBridge which was altered. The problem is that it is not clear if the tuning values for each CPU family is correct as the methodology used to tune the values is unclear. This patch uses a conservative tlb_flushall_shift value for all CPU families except IvyBridge so the decision can be revisited if any regression is found as a result of this change. IvyBridge is an exception as testing with one methodology determined that the value of 2 is acceptable. Details are in the changelog for the patch "x86: mm: Change tlb_flushall_shift for IvyBridge". One important aspect of this to watch out for is Xen. The original commit log mentioned large performance gains on Xen. It's possible Xen is more sensitive to this value if it flushes small ranges of pages more frequently than workloads on bare metal typically do. Signed-off-by: Mel Gorman <mgorman@suse.de> Tested-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Alex Shi <alex.shi@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/n/tip-dyzMww3fqugnhbhgo6Gxmtkw@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org> f98b7a77 Tue Jan 21 16:33:21 CST 2014 Mel Gorman <mgorman@suse.de> x86: mm: change tlb_flushall_shift for IvyBridge There was a large performance regression that was bisected to commit 611ae8e3 ("x86/tlb: enable tlb flush range support for x86"). This patch simply changes the default balance point between a local and global flush for IvyBridge. In the interest of allowing the tests to be reproduced, this patch was tested using mmtests 0.15 with the following configurations configs/config-global-dhp__tlbflush-performance configs/config-global-dhp__scheduler-performance configs/config-global-dhp__network-performance Results are from two machines Ivybridge 4 threads: Intel(R) Core(TM) i3-3240 CPU @ 3.40GHz Ivybridge 8 threads: Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz Page fault microbenchmark showed nothing interesting. Ebizzy was configured to run multiple iterations and threads. Thread counts ranged from 1 to NR_CPUS*2. For each thread count, it ran 100 iterations and each iteration lasted 10 seconds. Ivybridge 4 threads 3.13.0-rc7 3.13.0-rc7 vanilla altshift-v3 Mean 1 6395.44 ( 0.00%) 6789.09 ( 6.16%) Mean 2 7012.85 ( 0.00%) 8052.16 ( 14.82%) Mean 3 6403.04 ( 0.00%) 6973.74 ( 8.91%) Mean 4 6135.32 ( 0.00%) 6582.33 ( 7.29%) Mean 5 6095.69 ( 0.00%) 6526.68 ( 7.07%) Mean 6 6114.33 ( 0.00%) 6416.64 ( 4.94%) Mean 7 6085.10 ( 0.00%) 6448.51 ( 5.97%) Mean 8 6120.62 ( 0.00%) 6462.97 ( 5.59%) Ivybridge 8 threads 3.13.0-rc7 3.13.0-rc7 vanilla altshift-v3 Mean 1 7336.65 ( 0.00%) 7787.02 ( 6.14%) Mean 2 8218.41 ( 0.00%) 9484.13 ( 15.40%) Mean 3 7973.62 ( 0.00%) 8922.01 ( 11.89%) Mean 4 7798.33 ( 0.00%) 8567.03 ( 9.86%) Mean 5 7158.72 ( 0.00%) 8214.23 ( 14.74%) Mean 6 6852.27 ( 0.00%) 7952.45 ( 16.06%) Mean 7 6774.65 ( 0.00%) 7536.35 ( 11.24%) Mean 8 6510.50 ( 0.00%) 6894.05 ( 5.89%) Mean 12 6182.90 ( 0.00%) 6661.29 ( 7.74%) Mean 16 6100.09 ( 0.00%) 6608.69 ( 8.34%) Ebizzy hits the worst case scenario for TLB range flushing every time and it shows for these Ivybridge CPUs at least that the default choice is a poor on. The patch addresses the problem. Next was a tlbflush microbenchmark written by Alex Shi at http://marc.info/?l=linux-kernel&m=133727348217113 . It measures access costs while the TLB is being flushed. The expectation is that if there are always full TLB flushes that the benchmark would suffer and it benefits from range flushing There are 320 iterations of the test per thread count. The number of entries is randomly selected with a min of 1 and max of 512. To ensure a reasonably even spread of entries, the full range is broken up into 8 sections and a random number selected within that section. iteration 1, random number between 0-64 iteration 2, random number between 64-128 etc This is still a very weak methodology. When you do not know what are typical ranges, random is a reasonable choice but it can be easily argued that the opimisation was for smaller ranges and an even spread is not representative of any workload that matters. To improve this, we'd need to know the probability distribution of TLB flush range sizes for a set of workloads that are considered "common", build a synthetic trace and feed that into this benchmark. Even that is not perfect because it would not account for the time between flushes but there are limits of what can be reasonably done and still be doing something useful. If a representative synthetic trace is provided then this benchmark could be revisited and the shift values retuned. Ivybridge 4 threads 3.13.0-rc7 3.13.0-rc7 vanilla altshift-v3 Mean 1 10.50 ( 0.00%) 10.50 ( 0.03%) Mean 2 17.59 ( 0.00%) 17.18 ( 2.34%) Mean 3 22.98 ( 0.00%) 21.74 ( 5.41%) Mean 5 47.13 ( 0.00%) 46.23 ( 1.92%) Mean 8 43.30 ( 0.00%) 42.56 ( 1.72%) Ivybridge 8 threads 3.13.0-rc7 3.13.0-rc7 vanilla altshift-v3 Mean 1 9.45 ( 0.00%) 9.36 ( 0.93%) Mean 2 9.37 ( 0.00%) 9.70 ( -3.54%) Mean 3 9.36 ( 0.00%) 9.29 ( 0.70%) Mean 5 14.49 ( 0.00%) 15.04 ( -3.75%) Mean 8 41.08 ( 0.00%) 38.73 ( 5.71%) Mean 13 32.04 ( 0.00%) 31.24 ( 2.49%) Mean 16 40.05 ( 0.00%) 39.04 ( 2.51%) For both CPUs, average access time is reduced which is good as this is the benchmark that was used to tune the shift values in the first place albeit it is now known *how* the benchmark was used. The scheduler benchmarks were somewhat inconclusive. They showed gains and losses and makes me reconsider how stable those benchmarks really are or if something else might be interfering with the test results recently. Network benchmarks were inconclusive. Almost all results were flat except for netperf-udp tests on the 4 thread machine. These results were unstable and showed large variations between reboots. It is unknown if this is a recent problems but I've noticed before that netperf-udp results tend to vary. Based on these results, changing the default for Ivybridge seems like a logical choice. Signed-off-by: Mel Gorman <mgorman@suse.de> Tested-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Alex Shi <alex.shi@linaro.org> Reviewed-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/n/tip-cqnadffh1tiqrshthRj3Esge@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
|
H A D | amd.c | b9a3b4c9 Tue Jan 21 16:33:22 CST 2014 Mel Gorman <mgorman@suse.de> mm, x86: Revisit tlb_flushall_shift tuning for page flushes except on IvyBridge
There was a large ebizzy performance regression that was bisected to commit 611ae8e3 (x86/tlb: enable tlb flush range support for x86). The problem was related to the tlb_flushall_shift tuning for IvyBridge which was altered. The problem is that it is not clear if the tuning values for each CPU family is correct as the methodology used to tune the values is unclear.
This patch uses a conservative tlb_flushall_shift value for all CPU families except IvyBridge so the decision can be revisited if any regression is found as a result of this change. IvyBridge is an exception as testing with one methodology determined that the value of 2 is acceptable. Details are in the changelog for the patch "x86: mm: Change tlb_flushall_shift for IvyBridge".
One important aspect of this to watch out for is Xen. The original commit log mentioned large performance gains on Xen. It's possible Xen is more sensitive to this value if it flushes small ranges of pages more frequently than workloads on bare metal typically do.
Signed-off-by: Mel Gorman <mgorman@suse.de> Tested-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Alex Shi <alex.shi@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/n/tip-dyzMww3fqugnhbhgo6Gxmtkw@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org> b9a3b4c9 Tue Jan 21 16:33:22 CST 2014 Mel Gorman <mgorman@suse.de> mm, x86: Revisit tlb_flushall_shift tuning for page flushes except on IvyBridge There was a large ebizzy performance regression that was bisected to commit 611ae8e3 (x86/tlb: enable tlb flush range support for x86). The problem was related to the tlb_flushall_shift tuning for IvyBridge which was altered. The problem is that it is not clear if the tuning values for each CPU family is correct as the methodology used to tune the values is unclear. This patch uses a conservative tlb_flushall_shift value for all CPU families except IvyBridge so the decision can be revisited if any regression is found as a result of this change. IvyBridge is an exception as testing with one methodology determined that the value of 2 is acceptable. Details are in the changelog for the patch "x86: mm: Change tlb_flushall_shift for IvyBridge". One important aspect of this to watch out for is Xen. The original commit log mentioned large performance gains on Xen. It's possible Xen is more sensitive to this value if it flushes small ranges of pages more frequently than workloads on bare metal typically do. Signed-off-by: Mel Gorman <mgorman@suse.de> Tested-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Alex Shi <alex.shi@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/n/tip-dyzMww3fqugnhbhgo6Gxmtkw@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
|
/openbmc/linux/arch/x86/mm/ |
H A D | tlb.c | 611ae8e3 Wed Jun 27 20:02:22 CDT 2012 Alex Shi <alex.shi@intel.com> x86/tlb: enable tlb flush range support for x86
Not every tlb_flush execution moment is really need to evacuate all TLB entries, like in munmap, just few 'invlpg' is better for whole process performance, since it leaves most of TLB entries for later accessing.
This patch also rewrite flush_tlb_range for 2 purposes: 1, split it out to get flush_blt_mm_range function. 2, clean up to reduce line breaking, thanks for Borislav's input.
My micro benchmark 'mummap' http://lkml.org/lkml/2012/5/17/59 show that the random memory access on other CPU has 0~50% speed up on a 2P * 4cores * HT NHM EP while do 'munmap'.
Thanks Yongjie's testing on this patch: ------------- I used Linux 3.4-RC6 w/ and w/o his patches as Xen dom0 and guest kernel. After running two benchmarks in Xen HVM guest, I found his patches brought about 1%~3% performance gain in 'kernel build' and 'netperf' testing, though the performance gain was not very stable in 'kernel build' testing.
Some detailed testing results are below.
Testing Environment: Hardware: Romley-EP platform Xen version: latest upstream Linux kernel: 3.4-RC6 Guest vCPU number: 8 NIC: Intel 82599 (10GB bandwidth)
In 'kernel build' testing in guest: Command line | performance gain make -j 4 | 3.81% make -j 8 | 0.37% make -j 16 | -0.52%
In 'netperf' testing, we tested TCP_STREAM with default socket size 16384 byte as large packet and 64 byte as small packet. I used several clients to add networking pressure, then 'netperf' server automatically generated several threads to response them. I also used large-size packet and small-size packet in the testing. Packet size | Thread number | performance gain 16384 bytes | 4 | 0.02% 16384 bytes | 8 | 2.21% 16384 bytes | 16 | 2.04% 64 bytes | 4 | 1.07% 64 bytes | 8 | 3.31% 64 bytes | 16 | 0.71%
Signed-off-by: Alex Shi <alex.shi@intel.com> Link: http://lkml.kernel.org/r/1340845344-27557-8-git-send-email-alex.shi@intel.com Tested-by: Ren, Yongjie <yongjie.ren@intel.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com> 611ae8e3 Wed Jun 27 20:02:22 CDT 2012 Alex Shi <alex.shi@intel.com> x86/tlb: enable tlb flush range support for x86 Not every tlb_flush execution moment is really need to evacuate all TLB entries, like in munmap, just few 'invlpg' is better for whole process performance, since it leaves most of TLB entries for later accessing. This patch also rewrite flush_tlb_range for 2 purposes: 1, split it out to get flush_blt_mm_range function. 2, clean up to reduce line breaking, thanks for Borislav's input. My micro benchmark 'mummap' http://lkml.org/lkml/2012/5/17/59 show that the random memory access on other CPU has 0~50% speed up on a 2P * 4cores * HT NHM EP while do 'munmap'. Thanks Yongjie's testing on this patch: ------------- I used Linux 3.4-RC6 w/ and w/o his patches as Xen dom0 and guest kernel. After running two benchmarks in Xen HVM guest, I found his patches brought about 1%~3% performance gain in 'kernel build' and 'netperf' testing, though the performance gain was not very stable in 'kernel build' testing. Some detailed testing results are below. Testing Environment: Hardware: Romley-EP platform Xen version: latest upstream Linux kernel: 3.4-RC6 Guest vCPU number: 8 NIC: Intel 82599 (10GB bandwidth) In 'kernel build' testing in guest: Command line | performance gain make -j 4 | 3.81% make -j 8 | 0.37% make -j 16 | -0.52% In 'netperf' testing, we tested TCP_STREAM with default socket size 16384 byte as large packet and 64 byte as small packet. I used several clients to add networking pressure, then 'netperf' server automatically generated several threads to response them. I also used large-size packet and small-size packet in the testing. Packet size | Thread number | performance gain 16384 bytes | 4 | 0.02% 16384 bytes | 8 | 2.21% 16384 bytes | 16 | 2.04% 64 bytes | 4 | 1.07% 64 bytes | 8 | 3.31% 64 bytes | 16 | 0.71% Signed-off-by: Alex Shi <alex.shi@intel.com> Link: http://lkml.kernel.org/r/1340845344-27557-8-git-send-email-alex.shi@intel.com Tested-by: Ren, Yongjie <yongjie.ren@intel.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
|