Message ID | 20190624025604.30896-1-ying.huang@intel.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [-mm] autonuma: Fix scan period updating | expand |
On Mon, Jun 24, 2019 at 10:56:04AM +0800, Huang Ying wrote: > The autonuma scan period should be increased (scanning is slowed down) > if the majority of the page accesses are shared with other processes. > But in current code, the scan period will be decreased (scanning is > speeded up) in that situation. > > This patch fixes the code. And this has been tested via tracing the > scan period changing and /proc/vmstat numa_pte_updates counter when > running a multi-threaded memory accessing program (most memory > areas are accessed by multiple threads). > The patch somewhat flips the logic on whether shared or private is considered and it's not immediately obvious why that was required. That aside, other than the impact on numa_pte_updates, what actual performance difference was measured and on on what workloads?
On Mon, Jun 24, 2019 at 10:25 PM Mel Gorman <mgorman@suse.de> wrote: > > On Mon, Jun 24, 2019 at 10:56:04AM +0800, Huang Ying wrote: > > The autonuma scan period should be increased (scanning is slowed down) > > if the majority of the page accesses are shared with other processes. > > But in current code, the scan period will be decreased (scanning is > > speeded up) in that situation. > > > > This patch fixes the code. And this has been tested via tracing the > > scan period changing and /proc/vmstat numa_pte_updates counter when > > running a multi-threaded memory accessing program (most memory > > areas are accessed by multiple threads). > > > > The patch somewhat flips the logic on whether shared or private is > considered and it's not immediately obvious why that was required. That > aside, other than the impact on numa_pte_updates, what actual > performance difference was measured and on on what workloads? The original scanning period updating logic doesn't match the original patch description and comments. I think the original patch description and comments make more sense. So I fix the code logic to make it match the original patch description and comments. If my understanding to the original code logic and the original patch description and comments were correct, do you think the original patch description and comments are wrong so we need to fix the comments instead? Or you think we should prove whether the original patch description and comments are correct? Best Regards, Huang, Ying
On Tue, Jun 25, 2019 at 09:23:22PM +0800, huang ying wrote: > On Mon, Jun 24, 2019 at 10:25 PM Mel Gorman <mgorman@suse.de> wrote: > > > > On Mon, Jun 24, 2019 at 10:56:04AM +0800, Huang Ying wrote: > > > The autonuma scan period should be increased (scanning is slowed down) > > > if the majority of the page accesses are shared with other processes. > > > But in current code, the scan period will be decreased (scanning is > > > speeded up) in that situation. > > > > > > This patch fixes the code. And this has been tested via tracing the > > > scan period changing and /proc/vmstat numa_pte_updates counter when > > > running a multi-threaded memory accessing program (most memory > > > areas are accessed by multiple threads). > > > > > > > The patch somewhat flips the logic on whether shared or private is > > considered and it's not immediately obvious why that was required. That > > aside, other than the impact on numa_pte_updates, what actual > > performance difference was measured and on on what workloads? > > The original scanning period updating logic doesn't match the original > patch description and comments. I think the original patch > description and comments make more sense. So I fix the code logic to > make it match the original patch description and comments. > > If my understanding to the original code logic and the original patch > description and comments were correct, do you think the original patch > description and comments are wrong so we need to fix the comments > instead? Or you think we should prove whether the original patch > description and comments are correct? > I'm about to get knocked offline so cannot answer properly. The code may indeed be wrong and I have observed higher than expected NUMA scanning behaviour than expected although not enough to cause problems. A comment fix is fine but if you're changing the scanning behaviour, it should be backed up with data justifying that the change both reduces the observed scanning and that it has no adverse performance implications.
Mel Gorman <mgorman@suse.de> writes: > On Tue, Jun 25, 2019 at 09:23:22PM +0800, huang ying wrote: >> On Mon, Jun 24, 2019 at 10:25 PM Mel Gorman <mgorman@suse.de> wrote: >> > >> > On Mon, Jun 24, 2019 at 10:56:04AM +0800, Huang Ying wrote: >> > > The autonuma scan period should be increased (scanning is slowed down) >> > > if the majority of the page accesses are shared with other processes. >> > > But in current code, the scan period will be decreased (scanning is >> > > speeded up) in that situation. >> > > >> > > This patch fixes the code. And this has been tested via tracing the >> > > scan period changing and /proc/vmstat numa_pte_updates counter when >> > > running a multi-threaded memory accessing program (most memory >> > > areas are accessed by multiple threads). >> > > >> > >> > The patch somewhat flips the logic on whether shared or private is >> > considered and it's not immediately obvious why that was required. That >> > aside, other than the impact on numa_pte_updates, what actual >> > performance difference was measured and on on what workloads? >> >> The original scanning period updating logic doesn't match the original >> patch description and comments. I think the original patch >> description and comments make more sense. So I fix the code logic to >> make it match the original patch description and comments. >> >> If my understanding to the original code logic and the original patch >> description and comments were correct, do you think the original patch >> description and comments are wrong so we need to fix the comments >> instead? Or you think we should prove whether the original patch >> description and comments are correct? >> > > I'm about to get knocked offline so cannot answer properly. The code may > indeed be wrong and I have observed higher than expected NUMA scanning > behaviour than expected although not enough to cause problems. A comment > fix is fine but if you're changing the scanning behaviour, it should be > backed up with data justifying that the change both reduces the observed > scanning and that it has no adverse performance implications. Got it! Thanks for comments! As for performance testing, do you have some candidate workloads? Best Regards, Huang, Ying
On Thu, Jul 04, 2019 at 08:32:06AM +0800, Huang, Ying wrote: > Mel Gorman <mgorman@suse.de> writes: > > > On Tue, Jun 25, 2019 at 09:23:22PM +0800, huang ying wrote: > >> On Mon, Jun 24, 2019 at 10:25 PM Mel Gorman <mgorman@suse.de> wrote: > >> > > >> > On Mon, Jun 24, 2019 at 10:56:04AM +0800, Huang Ying wrote: > >> > > The autonuma scan period should be increased (scanning is slowed down) > >> > > if the majority of the page accesses are shared with other processes. > >> > > But in current code, the scan period will be decreased (scanning is > >> > > speeded up) in that situation. > >> > > > >> > > This patch fixes the code. And this has been tested via tracing the > >> > > scan period changing and /proc/vmstat numa_pte_updates counter when > >> > > running a multi-threaded memory accessing program (most memory > >> > > areas are accessed by multiple threads). > >> > > > >> > > >> > The patch somewhat flips the logic on whether shared or private is > >> > considered and it's not immediately obvious why that was required. That > >> > aside, other than the impact on numa_pte_updates, what actual > >> > performance difference was measured and on on what workloads? > >> > >> The original scanning period updating logic doesn't match the original > >> patch description and comments. I think the original patch > >> description and comments make more sense. So I fix the code logic to > >> make it match the original patch description and comments. > >> > >> If my understanding to the original code logic and the original patch > >> description and comments were correct, do you think the original patch > >> description and comments are wrong so we need to fix the comments > >> instead? Or you think we should prove whether the original patch > >> description and comments are correct? > >> > > > > I'm about to get knocked offline so cannot answer properly. The code may > > indeed be wrong and I have observed higher than expected NUMA scanning > > behaviour than expected although not enough to cause problems. A comment > > fix is fine but if you're changing the scanning behaviour, it should be > > backed up with data justifying that the change both reduces the observed > > scanning and that it has no adverse performance implications. > > Got it! Thanks for comments! As for performance testing, do you have > some candidate workloads? > Ordinarily I would hope that the patch was motivated by observed behaviour so you have a metric for goodness. However, for NUMA balancing I would typically run basic workloads first -- dbench, tbench, netperf, hackbench and pipetest. The objective would be to measure the degree automatic NUMA balancing is interfering with a basic workload to see if they patch reduces the number of minor faults incurred even though there is no NUMA balancing to be worried about. This measures the general overhead of a patch. If your reasoning is correct, you'd expect lower overhead. For balancing itself, I usually look at Andrea's original autonuma benchmark, NAS Parallel Benchmark (D class usually although C class for much older or smaller machines) and spec JBB 2005 and 2015. Of the JBB benchmarks, 2005 is usually more reasonable for evaluating NUMA balancing than 2015 is (which can be unstable for a variety of reasons). In this case, I would be looking at whether the overhead is reduced, whether the ratio of local hits is the same or improved and the primary metric of each (time to completion for Andrea's and NAS, throughput for JBB). Even if there is no change to locality and the primary metric but there is less scanning and overhead overall, it would still be an improvement. If you have trouble doing such an evaluation, I'll queue tests if they are based on a patch that addresses the specific point of concern (scan period not updated) as it's still not obvious why flipping the logic of whether shared or private is considered was necessary.
Mel Gorman <mgorman@suse.de> writes: > On Thu, Jul 04, 2019 at 08:32:06AM +0800, Huang, Ying wrote: >> Mel Gorman <mgorman@suse.de> writes: >> >> > On Tue, Jun 25, 2019 at 09:23:22PM +0800, huang ying wrote: >> >> On Mon, Jun 24, 2019 at 10:25 PM Mel Gorman <mgorman@suse.de> wrote: >> >> > >> >> > On Mon, Jun 24, 2019 at 10:56:04AM +0800, Huang Ying wrote: >> >> > > The autonuma scan period should be increased (scanning is slowed down) >> >> > > if the majority of the page accesses are shared with other processes. >> >> > > But in current code, the scan period will be decreased (scanning is >> >> > > speeded up) in that situation. >> >> > > >> >> > > This patch fixes the code. And this has been tested via tracing the >> >> > > scan period changing and /proc/vmstat numa_pte_updates counter when >> >> > > running a multi-threaded memory accessing program (most memory >> >> > > areas are accessed by multiple threads). >> >> > > >> >> > >> >> > The patch somewhat flips the logic on whether shared or private is >> >> > considered and it's not immediately obvious why that was required. That >> >> > aside, other than the impact on numa_pte_updates, what actual >> >> > performance difference was measured and on on what workloads? >> >> >> >> The original scanning period updating logic doesn't match the original >> >> patch description and comments. I think the original patch >> >> description and comments make more sense. So I fix the code logic to >> >> make it match the original patch description and comments. >> >> >> >> If my understanding to the original code logic and the original patch >> >> description and comments were correct, do you think the original patch >> >> description and comments are wrong so we need to fix the comments >> >> instead? Or you think we should prove whether the original patch >> >> description and comments are correct? >> >> >> > >> > I'm about to get knocked offline so cannot answer properly. The code may >> > indeed be wrong and I have observed higher than expected NUMA scanning >> > behaviour than expected although not enough to cause problems. A comment >> > fix is fine but if you're changing the scanning behaviour, it should be >> > backed up with data justifying that the change both reduces the observed >> > scanning and that it has no adverse performance implications. >> >> Got it! Thanks for comments! As for performance testing, do you have >> some candidate workloads? >> > > Ordinarily I would hope that the patch was motivated by observed > behaviour so you have a metric for goodness. However, for NUMA balancing > I would typically run basic workloads first -- dbench, tbench, netperf, > hackbench and pipetest. The objective would be to measure the degree > automatic NUMA balancing is interfering with a basic workload to see if > they patch reduces the number of minor faults incurred even though there > is no NUMA balancing to be worried about. This measures the general > overhead of a patch. If your reasoning is correct, you'd expect lower > overhead. > > For balancing itself, I usually look at Andrea's original autonuma > benchmark, NAS Parallel Benchmark (D class usually although C class for > much older or smaller machines) and spec JBB 2005 and 2015. Of the JBB > benchmarks, 2005 is usually more reasonable for evaluating NUMA balancing > than 2015 is (which can be unstable for a variety of reasons). In this > case, I would be looking at whether the overhead is reduced, whether the > ratio of local hits is the same or improved and the primary metric of > each (time to completion for Andrea's and NAS, throughput for JBB). > > Even if there is no change to locality and the primary metric but there > is less scanning and overhead overall, it would still be an improvement. Thanks a lot for your detailed guidance. > If you have trouble doing such an evaluation, I'll queue tests if they > are based on a patch that addresses the specific point of concern (scan > period not updated) as it's still not obvious why flipping the logic of > whether shared or private is considered was necessary. I can do the evaluation, but it will take quite some time for me to setup and run all these benchmarks. So if these benchmarks have already been setup in your environment, so that your extra effort is minimal, it will be great if you can queue tests for the patch. Feel free to reject me for any inconvenience. Best Regards, Huang, Ying
On Fri, Jul 12, 2019 at 06:48:05PM +0800, Huang, Ying wrote: > > Ordinarily I would hope that the patch was motivated by observed > > behaviour so you have a metric for goodness. However, for NUMA balancing > > I would typically run basic workloads first -- dbench, tbench, netperf, > > hackbench and pipetest. The objective would be to measure the degree > > automatic NUMA balancing is interfering with a basic workload to see if > > they patch reduces the number of minor faults incurred even though there > > is no NUMA balancing to be worried about. This measures the general > > overhead of a patch. If your reasoning is correct, you'd expect lower > > overhead. > > > > For balancing itself, I usually look at Andrea's original autonuma > > benchmark, NAS Parallel Benchmark (D class usually although C class for > > much older or smaller machines) and spec JBB 2005 and 2015. Of the JBB > > benchmarks, 2005 is usually more reasonable for evaluating NUMA balancing > > than 2015 is (which can be unstable for a variety of reasons). In this > > case, I would be looking at whether the overhead is reduced, whether the > > ratio of local hits is the same or improved and the primary metric of > > each (time to completion for Andrea's and NAS, throughput for JBB). > > > > Even if there is no change to locality and the primary metric but there > > is less scanning and overhead overall, it would still be an improvement. > > Thanks a lot for your detailed guidance. > No problem. > > If you have trouble doing such an evaluation, I'll queue tests if they > > are based on a patch that addresses the specific point of concern (scan > > period not updated) as it's still not obvious why flipping the logic of > > whether shared or private is considered was necessary. > > I can do the evaluation, but it will take quite some time for me to > setup and run all these benchmarks. So if these benchmarks have already > been setup in your environment, so that your extra effort is minimal, it > will be great if you can queue tests for the patch. Feel free to reject > me for any inconvenience. > They're not setup as such, but my testing infrastructure is heavily automated so it's easy to do and I think it's worth looking at. If you update your patch to target just the scan period aspects, I'll queue it up and get back to you. It usually takes a few days for the automation to finish whatever it's doing and pick up a patch for evaluation.
Mel Gorman <mgorman@suse.de> writes: > On Fri, Jul 12, 2019 at 06:48:05PM +0800, Huang, Ying wrote: >> > Ordinarily I would hope that the patch was motivated by observed >> > behaviour so you have a metric for goodness. However, for NUMA balancing >> > I would typically run basic workloads first -- dbench, tbench, netperf, >> > hackbench and pipetest. The objective would be to measure the degree >> > automatic NUMA balancing is interfering with a basic workload to see if >> > they patch reduces the number of minor faults incurred even though there >> > is no NUMA balancing to be worried about. This measures the general >> > overhead of a patch. If your reasoning is correct, you'd expect lower >> > overhead. >> > >> > For balancing itself, I usually look at Andrea's original autonuma >> > benchmark, NAS Parallel Benchmark (D class usually although C class for >> > much older or smaller machines) and spec JBB 2005 and 2015. Of the JBB >> > benchmarks, 2005 is usually more reasonable for evaluating NUMA balancing >> > than 2015 is (which can be unstable for a variety of reasons). In this >> > case, I would be looking at whether the overhead is reduced, whether the >> > ratio of local hits is the same or improved and the primary metric of >> > each (time to completion for Andrea's and NAS, throughput for JBB). >> > >> > Even if there is no change to locality and the primary metric but there >> > is less scanning and overhead overall, it would still be an improvement. >> >> Thanks a lot for your detailed guidance. >> > > No problem. > >> > If you have trouble doing such an evaluation, I'll queue tests if they >> > are based on a patch that addresses the specific point of concern (scan >> > period not updated) as it's still not obvious why flipping the logic of >> > whether shared or private is considered was necessary. >> >> I can do the evaluation, but it will take quite some time for me to >> setup and run all these benchmarks. So if these benchmarks have already >> been setup in your environment, so that your extra effort is minimal, it >> will be great if you can queue tests for the patch. Feel free to reject >> me for any inconvenience. >> > > They're not setup as such, but my testing infrastructure is heavily > automated so it's easy to do and I think it's worth looking at. If you > update your patch to target just the scan period aspects, I'll queue it > up and get back to you. It usually takes a few days for the automation > to finish whatever it's doing and pick up a patch for evaluation. Thanks a lot for your help! The updated patch is as follows. It targets only the scan period aspects. Best Regards, Huang, Ying ----------------------8<---------------------------- From 910a52cbf5a521c1562a573904c9507d0367bb0f Mon Sep 17 00:00:00 2001 From: Huang Ying <ying.huang@intel.com> Date: Sat, 22 Jun 2019 17:36:29 +0800 Subject: [PATCH] autonuma: Fix scan period updating From the commit log and comments of commit 37ec97deb3a8 ("sched/numa: Slow down scan rate if shared faults dominate"), the autonuma scan period should be increased (scanning is slowed down) if the majority of the page accesses are shared with other processes. But in current code, the scan period will be decreased (scanning is speeded up) in that situation. The commit log and comments make more sense. So this patch fixes the code to make it match the commit log and comments. And this has been verified via tracing the scan period changing and /proc/vmstat numa_pte_updates counter when running a multi-threaded memory accessing program (most memory areas are accessed by multiple threads). Fixes: 37ec97deb3a8 ("sched/numa: Slow down scan rate if shared faults dominate") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Mel Gorman <mgorman@suse.de> Cc: jhladky@redhat.com Cc: lvenanci@redhat.com Cc: Ingo Molnar <mingo@kernel.org> --- kernel/sched/fair.c | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 036be95a87e9..468a1c5038b2 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -1940,7 +1940,7 @@ static void update_task_scan_period(struct task_struct *p, unsigned long shared, unsigned long private) { unsigned int period_slot; - int lr_ratio, ps_ratio; + int lr_ratio, sp_ratio; int diff; unsigned long remote = p->numa_faults_locality[0]; @@ -1971,22 +1971,22 @@ static void update_task_scan_period(struct task_struct *p, */ period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); - ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); + sp_ratio = (shared * NUMA_PERIOD_SLOTS) / (private + shared); - if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { + if (sp_ratio >= NUMA_PERIOD_THRESHOLD) { /* - * Most memory accesses are local. There is no need to - * do fast NUMA scanning, since memory is already local. + * Most memory accesses are shared with other tasks. + * There is no point in continuing fast NUMA scanning, + * since other tasks may just move the memory elsewhere. */ - int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; + int slot = sp_ratio - NUMA_PERIOD_THRESHOLD; if (!slot) slot = 1; diff = slot * period_slot; } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { /* - * Most memory accesses are shared with other tasks. - * There is no point in continuing fast NUMA scanning, - * since other tasks may just move the memory elsewhere. + * Most memory accesses are local. There is no need to + * do fast NUMA scanning, since memory is already local. */ int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; if (!slot) @@ -1998,7 +1998,7 @@ static void update_task_scan_period(struct task_struct *p, * yet they are not on the local NUMA node. Speed up * NUMA scanning to get the memory moved over. */ - int ratio = max(lr_ratio, ps_ratio); + int ratio = max(lr_ratio, sp_ratio); diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; }
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index f35930f5e528..79bc4d2d1e58 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -1923,7 +1923,7 @@ static void update_task_scan_period(struct task_struct *p, unsigned long shared, unsigned long private) { unsigned int period_slot; - int lr_ratio, ps_ratio; + int lr_ratio, sp_ratio; int diff; unsigned long remote = p->numa_faults_locality[0]; @@ -1954,22 +1954,22 @@ static void update_task_scan_period(struct task_struct *p, */ period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); - ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); + sp_ratio = (shared * NUMA_PERIOD_SLOTS) / (private + shared); - if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { + if (sp_ratio >= NUMA_PERIOD_THRESHOLD) { /* - * Most memory accesses are local. There is no need to - * do fast NUMA scanning, since memory is already local. + * Most memory accesses are shared with other tasks. + * There is no point in continuing fast NUMA scanning, + * since other tasks may just move the memory elsewhere. */ - int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; + int slot = sp_ratio - NUMA_PERIOD_THRESHOLD; if (!slot) slot = 1; diff = slot * period_slot; } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { /* - * Most memory accesses are shared with other tasks. - * There is no point in continuing fast NUMA scanning, - * since other tasks may just move the memory elsewhere. + * Most memory accesses are local. There is no need to + * do fast NUMA scanning, since memory is already local. */ int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; if (!slot) @@ -1981,7 +1981,7 @@ static void update_task_scan_period(struct task_struct *p, * yet they are not on the local NUMA node. Speed up * NUMA scanning to get the memory moved over. */ - int ratio = max(lr_ratio, ps_ratio); + int ratio = max(lr_ratio, sp_ratio); diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; }
The autonuma scan period should be increased (scanning is slowed down) if the majority of the page accesses are shared with other processes. But in current code, the scan period will be decreased (scanning is speeded up) in that situation. This patch fixes the code. And this has been tested via tracing the scan period changing and /proc/vmstat numa_pte_updates counter when running a multi-threaded memory accessing program (most memory areas are accessed by multiple threads). Fixes: 37ec97deb3a8 ("sched/numa: Slow down scan rate if shared faults dominate") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Mel Gorman <mgorman@suse.de> Cc: jhladky@redhat.com Cc: lvenanci@redhat.com Cc: Ingo Molnar <mingo@kernel.org> --- kernel/sched/fair.c | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-)