| Message ID | 20201028023411.15045-3-ying.huang@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | autonuma: Migrate on fault among multiple bound nodes | expand |
On Wed, Oct 28, 2020 at 10:34:11AM +0800, Huang Ying wrote: > Now, AutoNUMA can only optimize the page placement among the NUMA nodes if the > default memory policy is used. Because the memory policy specified explicitly > should take precedence. But this seems too strict in some situations. For > example, on a system with 4 NUMA nodes, if the memory of an application is bound > to the node 0 and 1, AutoNUMA can potentially migrate the pages between the node > 0 and 1 to reduce cross-node accessing without breaking the explicit memory > binding policy. > > So in this patch, if mbind(.mode=MPOL_BIND, .flags=MPOL_MF_LAZY) is used to bind > the memory of the application to multiple nodes, and in the hint page fault > handler both the faulting page node and the accessing node are in the policy > nodemask, the page will be tried to be migrated to the accessing node to reduce > the cross-node accessing. > > [Peter Zijlstra: provided the simplified implementation method.] > > Questions: > > Sysctl knob kernel.numa_balancing can enable/disable AutoNUMA optimizing > globally. But for the memory areas that are bound to multiple NUMA nodes, even > if the AutoNUMA is enabled globally via the sysctl knob, we still need to enable > AutoNUMA again with a special flag. Why not just optimize the page placement if > possible as long as AutoNUMA is enabled globally? The interface would look > simpler with that. > > Signed-off-by: "Huang, Ying" <ying.huang@intel.com> I've no specific objection to the patch or the name change. I can't remember exactly why I picked the name, it was 8 years ago but I think it was because the policy represented the most basic possible approach that could be done without any attempt at being intelligent and established a baseline. The intent was that anything built on top had to be better than the most basic policy imaginable. The name reflected the dictionary definition at the time and happened to match the acronym closely enough and I wanted to make it absolutely clear to reviewers that the policy was not good enough (ruling out MPOL_BASIC or variants thereof) even if it happened to work for some workload and there was no intent to report it to the userspace API. The only hazard with the patch is that applications that use MPOL_BIND on multiple nodes may now incur some NUMA balancing overhead due to trapping faults and migrations. It might still end up being better but I was not aware of a *realistic* workload that binds to multiple nodes deliberately. Generally I expect if an application is binding, it's binding to one local node. If it shows up in regressions, it'll be interesting to get a detailed description of the workload. Pay particular attention to if THP is disabled as I learned relatively recently that NUMA balancing with THP disabled has higher overhead (which is hardly surprising). Lacking data or a specific objection Acked-by: Mel Gorman <mgorman@suse.de>
Hi, Mel, Thanks for comments! Mel Gorman <mgorman@suse.de> writes: > On Wed, Oct 28, 2020 at 10:34:11AM +0800, Huang Ying wrote: >> Now, AutoNUMA can only optimize the page placement among the NUMA nodes if the >> default memory policy is used. Because the memory policy specified explicitly >> should take precedence. But this seems too strict in some situations. For >> example, on a system with 4 NUMA nodes, if the memory of an application is bound >> to the node 0 and 1, AutoNUMA can potentially migrate the pages between the node >> 0 and 1 to reduce cross-node accessing without breaking the explicit memory >> binding policy. >> >> So in this patch, if mbind(.mode=MPOL_BIND, .flags=MPOL_MF_LAZY) is used to bind >> the memory of the application to multiple nodes, and in the hint page fault >> handler both the faulting page node and the accessing node are in the policy >> nodemask, the page will be tried to be migrated to the accessing node to reduce >> the cross-node accessing. >> >> [Peter Zijlstra: provided the simplified implementation method.] >> >> Questions: >> >> Sysctl knob kernel.numa_balancing can enable/disable AutoNUMA optimizing >> globally. But for the memory areas that are bound to multiple NUMA nodes, even >> if the AutoNUMA is enabled globally via the sysctl knob, we still need to enable >> AutoNUMA again with a special flag. Why not just optimize the page placement if >> possible as long as AutoNUMA is enabled globally? The interface would look >> simpler with that. >> >> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> > > I've no specific objection to the patch or the name change. I can't > remember exactly why I picked the name, it was 8 years ago but I think it > was because the policy represented the most basic possible approach that > could be done without any attempt at being intelligent and established > a baseline. The intent was that anything built on top had to be better > than the most basic policy imaginable. The name reflected the dictionary > definition at the time and happened to match the acronym closely enough > and I wanted to make it absolutely clear to reviewers that the policy > was not good enough (ruling out MPOL_BASIC or variants thereof) even if > it happened to work for some workload and there was no intent to report > it to the userspace API. > > The only hazard with the patch is that applications that use MPOL_BIND > on multiple nodes may now incur some NUMA balancing overhead due to > trapping faults and migrations. For this specific version of patch, I don't think this will happen. Because now, MPOL_F_MOF need to be set in struct mempolicy, for MPOL_BIND, only if mbind() syscall is called with MPOL_MF_LAZY, that will be the case. So I think most workloads will not be affected by this patch. The feature is opt-in. But from another point of view, I suggest to remove the constraints of MPOL_F_MOF in the future. If the overhead of AutoNUMA isn't acceptable, why not just disable AutoNUMA globally via sysctl knob? > It might still end up being better but I was not aware of a > *realistic* workload that binds to multiple nodes > deliberately. Generally I expect if an application is binding, it's > binding to one local node. Yes. It's not popular configuration for now. But for the memory tiering system with both DRAM and PMEM, the DRAM and PMEM in one socket will become 2 NUMA nodes. To avoid too much cross-socket memory accessing, but take advantage of both the DRAM and PMEM, the workload can be bound to 2 NUMA nodes (DRAM and PMEM). > If it shows up in regressions, it'll be interesting to get a detailed > description of the workload. Pay particular attention to if THP is > disabled as I learned relatively recently that NUMA balancing with THP > disabled has higher overhead (which is hardly surprising). Got it! > Lacking data or a specific objection > > Acked-by: Mel Gorman <mgorman@suse.de> Thanks! Best Regards, Huang, Ying
On Wed, Nov 04, 2020 at 01:36:58PM +0800, Huang, Ying wrote: > > I've no specific objection to the patch or the name change. I can't > > remember exactly why I picked the name, it was 8 years ago but I think it > > was because the policy represented the most basic possible approach that > > could be done without any attempt at being intelligent and established > > a baseline. The intent was that anything built on top had to be better > > than the most basic policy imaginable. The name reflected the dictionary > > definition at the time and happened to match the acronym closely enough > > and I wanted to make it absolutely clear to reviewers that the policy > > was not good enough (ruling out MPOL_BASIC or variants thereof) even if > > it happened to work for some workload and there was no intent to report > > it to the userspace API. > > > > The only hazard with the patch is that applications that use MPOL_BIND > > on multiple nodes may now incur some NUMA balancing overhead due to > > trapping faults and migrations. > > For this specific version of patch, I don't think this will happen. > Because now, MPOL_F_MOF need to be set in struct mempolicy, for > MPOL_BIND, only if mbind() syscall is called with MPOL_MF_LAZY, that > will be the case. So I think most workloads will not be affected by > this patch. The feature is opt-in. > Ok. > But from another point of view, I suggest to remove the constraints of > MPOL_F_MOF in the future. If the overhead of AutoNUMA isn't acceptable, > why not just disable AutoNUMA globally via sysctl knob? > Because it's a double edged sword. NUMA Balancing can make a workload faster while still incurring more overhead than it should -- particularly when threads are involved rescanning the same or unrelated regions. Global disabling only really should happen when an application is running that is the only application on the machine and has full NUMA awareness. > > It might still end up being better but I was not aware of a > > *realistic* workload that binds to multiple nodes > > deliberately. Generally I expect if an application is binding, it's > > binding to one local node. > > Yes. It's not popular configuration for now. But for the memory > tiering system with both DRAM and PMEM, the DRAM and PMEM in one socket > will become 2 NUMA nodes. To avoid too much cross-socket memory > accessing, but take advantage of both the DRAM and PMEM, the workload > can be bound to 2 NUMA nodes (DRAM and PMEM). > Ok, that may lead to unpredictable performance as it'll have variable performance with limited control of the "important" applications that should use DRAM over PMEM. That's a long road but the step is not incompatible with the long-term goal.
Mel Gorman <mgorman@suse.de> writes: > On Wed, Nov 04, 2020 at 01:36:58PM +0800, Huang, Ying wrote: >> But from another point of view, I suggest to remove the constraints of >> MPOL_F_MOF in the future. If the overhead of AutoNUMA isn't acceptable, >> why not just disable AutoNUMA globally via sysctl knob? >> > > Because it's a double edged sword. NUMA Balancing can make a workload > faster while still incurring more overhead than it should -- particularly > when threads are involved rescanning the same or unrelated regions. > Global disabling only really should happen when an application is running > that is the only application on the machine and has full NUMA awareness. Got it. So NUMA Balancing may in generally benefit some workloads but hurt some other workloads on one machine. So we need a method to enable/disable NUMA Balancing for one workload. Previously, this is done via the explicit NUMA policy. If some explicit NUMA policy is specified, NUMA Balancing is disabled for the memory region or the thread. And this can be reverted again for a memory region via MPOL_MF_LAZY. It appears that we lacks MPOL_MF_LAZY for the thread yet. >> > It might still end up being better but I was not aware of a >> > *realistic* workload that binds to multiple nodes >> > deliberately. Generally I expect if an application is binding, it's >> > binding to one local node. >> >> Yes. It's not popular configuration for now. But for the memory >> tiering system with both DRAM and PMEM, the DRAM and PMEM in one socket >> will become 2 NUMA nodes. To avoid too much cross-socket memory >> accessing, but take advantage of both the DRAM and PMEM, the workload >> can be bound to 2 NUMA nodes (DRAM and PMEM). >> > > Ok, that may lead to unpredictable performance as it'll have variable > performance with limited control of the "important" applications that > should use DRAM over PMEM. That's a long road but the step is not > incompatible with the long-term goal. Yes. Ben Widawsky is working on a patchset to make it possible to prefer the remote DRAM instead of the local PMEM as follows, https://lore.kernel.org/linux-mm/20200630212517.308045-1-ben.widawsky@intel.com/ Best Regards, Huang, Ying
On 20-11-06 15:28:59, Huang, Ying wrote: > Mel Gorman <mgorman@suse.de> writes: > > > On Wed, Nov 04, 2020 at 01:36:58PM +0800, Huang, Ying wrote: > >> But from another point of view, I suggest to remove the constraints of > >> MPOL_F_MOF in the future. If the overhead of AutoNUMA isn't acceptable, > >> why not just disable AutoNUMA globally via sysctl knob? > >> > > > > Because it's a double edged sword. NUMA Balancing can make a workload > > faster while still incurring more overhead than it should -- particularly > > when threads are involved rescanning the same or unrelated regions. > > Global disabling only really should happen when an application is running > > that is the only application on the machine and has full NUMA awareness. > > Got it. So NUMA Balancing may in generally benefit some workloads but > hurt some other workloads on one machine. So we need a method to > enable/disable NUMA Balancing for one workload. Previously, this is > done via the explicit NUMA policy. If some explicit NUMA policy is > specified, NUMA Balancing is disabled for the memory region or the > thread. And this can be reverted again for a memory region via > MPOL_MF_LAZY. It appears that we lacks MPOL_MF_LAZY for the thread yet. > > >> > It might still end up being better but I was not aware of a > >> > *realistic* workload that binds to multiple nodes > >> > deliberately. Generally I expect if an application is binding, it's > >> > binding to one local node. > >> > >> Yes. It's not popular configuration for now. But for the memory > >> tiering system with both DRAM and PMEM, the DRAM and PMEM in one socket > >> will become 2 NUMA nodes. To avoid too much cross-socket memory > >> accessing, but take advantage of both the DRAM and PMEM, the workload > >> can be bound to 2 NUMA nodes (DRAM and PMEM). > >> > > > > Ok, that may lead to unpredictable performance as it'll have variable > > performance with limited control of the "important" applications that > > should use DRAM over PMEM. That's a long road but the step is not > > incompatible with the long-term goal. > > Yes. Ben Widawsky is working on a patchset to make it possible to > prefer the remote DRAM instead of the local PMEM as follows, > > https://lore.kernel.org/linux-mm/20200630212517.308045-1-ben.widawsky@intel.com/ > > Best Regards, > Huang, Ying > Rebased version was posted here: https://lore.kernel.org/linux-mm/20201030190238.306764-1-ben.widawsky@intel.com/ Thanks. Ben
Hi, Mel, Mel Gorman <mgorman@suse.de> writes: > On Wed, Nov 04, 2020 at 01:36:58PM +0800, Huang, Ying wrote: >> > I've no specific objection to the patch or the name change. I can't >> > remember exactly why I picked the name, it was 8 years ago but I think it >> > was because the policy represented the most basic possible approach that >> > could be done without any attempt at being intelligent and established >> > a baseline. The intent was that anything built on top had to be better >> > than the most basic policy imaginable. The name reflected the dictionary >> > definition at the time and happened to match the acronym closely enough >> > and I wanted to make it absolutely clear to reviewers that the policy >> > was not good enough (ruling out MPOL_BASIC or variants thereof) even if >> > it happened to work for some workload and there was no intent to report >> > it to the userspace API. >> > >> > The only hazard with the patch is that applications that use MPOL_BIND >> > on multiple nodes may now incur some NUMA balancing overhead due to >> > trapping faults and migrations. >> >> For this specific version of patch, I don't think this will happen. >> Because now, MPOL_F_MOF need to be set in struct mempolicy, for >> MPOL_BIND, only if mbind() syscall is called with MPOL_MF_LAZY, that >> will be the case. So I think most workloads will not be affected by >> this patch. The feature is opt-in. >> > > Ok. I just found MPOL_MF_LAZY is disabled now. And as in commit a720094ded8c ("mm: mempolicy: Hide MPOL_NOOP and MPOL_MF_LAZY from userspace for now"), the ABI needs to be revisted before exporting to the user space formally. Sorry about that, I should have found that earlier. Think about that. I think MPOL_MF_LAZY is tied with MPOL_MF_MOVE, so it's semantics isn't good for the purpose of the patch. So I have rewritten the patch and the description and sent it as follows, can you help to review it? https://lore.kernel.org/lkml/20201111063717.186589-1-ying.huang@intel.com/ Best Regards, Huang, Ying
diff --git a/mm/mempolicy.c b/mm/mempolicy.c index f6948b659643..d0d25c2601a0 100644 --- a/mm/mempolicy.c +++ b/mm/mempolicy.c @@ -2490,15 +2490,19 @@ int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long break; case MPOL_BIND: - /* - * allows binding to multiple nodes. - * use current page if in policy nodemask, - * else select nearest allowed node, if any. - * If no allowed nodes, use current [!misplaced]. + * Allows binding to multiple nodes. If both current and + * accessing nodes are in policy nodemask, migrate to + * accessing node to optimize page placement. Otherwise, + * use current page if in policy nodemask, else select + * nearest allowed node, if any. If no allowed nodes, use + * current [!misplaced]. */ - if (node_isset(curnid, pol->v.nodes)) + if (node_isset(curnid, pol->v.nodes)) { + if (node_isset(thisnid, pol->v.nodes)) + goto mopron; goto out; + } z = first_zones_zonelist( node_zonelist(numa_node_id(), GFP_HIGHUSER), gfp_zone(GFP_HIGHUSER), @@ -2512,6 +2516,7 @@ int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long /* Migrate the page towards the node whose CPU is referencing it */ if (pol->flags & MPOL_F_MOPRON) { +mopron: polnid = thisnid; if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
Now, AutoNUMA can only optimize the page placement among the NUMA nodes if the default memory policy is used. Because the memory policy specified explicitly should take precedence. But this seems too strict in some situations. For example, on a system with 4 NUMA nodes, if the memory of an application is bound to the node 0 and 1, AutoNUMA can potentially migrate the pages between the node 0 and 1 to reduce cross-node accessing without breaking the explicit memory binding policy. So in this patch, if mbind(.mode=MPOL_BIND, .flags=MPOL_MF_LAZY) is used to bind the memory of the application to multiple nodes, and in the hint page fault handler both the faulting page node and the accessing node are in the policy nodemask, the page will be tried to be migrated to the accessing node to reduce the cross-node accessing. [Peter Zijlstra: provided the simplified implementation method.] Questions: Sysctl knob kernel.numa_balancing can enable/disable AutoNUMA optimizing globally. But for the memory areas that are bound to multiple NUMA nodes, even if the AutoNUMA is enabled globally via the sysctl knob, we still need to enable AutoNUMA again with a special flag. Why not just optimize the page placement if possible as long as AutoNUMA is enabled globally? The interface would look simpler with that. Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Mel Gorman <mgorman@suse.de> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: David Rientjes <rientjes@google.com> --- mm/mempolicy.c | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-)