| Message ID | 20200504070304.127361-1-sandipan@linux.ibm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | mm: vmstat: Use zeroed stats for unpopulated zones | expand |
On Mon 04-05-20 12:33:04, Sandipan Das wrote: > For unpopulated zones, the pagesets point to the common > boot_pageset which can have non-zero vm_numa_stat counts. > Because of this memory-less nodes end up having non-zero > NUMA statistics. This can be observed on any architecture > that supports memory-less NUMA nodes. > > E.g. > > $ numactl -H > available: 2 nodes (0-1) > node 0 cpus: 0 1 2 3 > node 0 size: 0 MB > node 0 free: 0 MB > node 1 cpus: 4 5 6 7 > node 1 size: 8131 MB > node 1 free: 6980 MB > node distances: > node 0 1 > 0: 10 40 > 1: 40 10 > > $ numastat > node0 node1 > numa_hit 108 56495 > numa_miss 0 0 > numa_foreign 0 0 > interleave_hit 0 4537 > local_node 108 31547 > other_node 0 24948 > > Hence, return zero explicitly for all the stats of an > unpopulated zone. I hope I am not just confused but I would expect that at least numa_foreign and other_node to be non zero.
On 5/4/20 12:26 PM, Michal Hocko wrote: > On Mon 04-05-20 12:33:04, Sandipan Das wrote: >> For unpopulated zones, the pagesets point to the common >> boot_pageset which can have non-zero vm_numa_stat counts. >> Because of this memory-less nodes end up having non-zero >> NUMA statistics. This can be observed on any architecture >> that supports memory-less NUMA nodes. >> >> E.g. >> >> $ numactl -H >> available: 2 nodes (0-1) >> node 0 cpus: 0 1 2 3 >> node 0 size: 0 MB >> node 0 free: 0 MB >> node 1 cpus: 4 5 6 7 >> node 1 size: 8131 MB >> node 1 free: 6980 MB >> node distances: >> node 0 1 >> 0: 10 40 >> 1: 40 10 >> >> $ numastat >> node0 node1 >> numa_hit 108 56495 >> numa_miss 0 0 >> numa_foreign 0 0 >> interleave_hit 0 4537 >> local_node 108 31547 >> other_node 0 24948 >> >> Hence, return zero explicitly for all the stats of an >> unpopulated zone. > > I hope I am not just confused but I would expect that at least > numa_foreign and other_node to be non zero. Hmm, checking zone_statistics(): NUMA_FOREIGN increment uses preferred zone, which is the first in zone in zonelist, so it will be a zone from node 1 even for allocations on cpu associated to node 0 - assuming node 0's unpopulated zones are not included in node 0's zonelist. NUMA_OTHER uses numa_node_id(), which would mean the node 0's cpus have node 1 in their numa_node_id() ? Is that correct? So the 108 comes from where exactly, some early allocations until all was initialized?
On Wed 06-05-20 15:33:36, Vlastimil Babka wrote: > On 5/4/20 12:26 PM, Michal Hocko wrote: > > On Mon 04-05-20 12:33:04, Sandipan Das wrote: > >> For unpopulated zones, the pagesets point to the common > >> boot_pageset which can have non-zero vm_numa_stat counts. > >> Because of this memory-less nodes end up having non-zero > >> NUMA statistics. This can be observed on any architecture > >> that supports memory-less NUMA nodes. > >> > >> E.g. > >> > >> $ numactl -H > >> available: 2 nodes (0-1) > >> node 0 cpus: 0 1 2 3 > >> node 0 size: 0 MB > >> node 0 free: 0 MB > >> node 1 cpus: 4 5 6 7 > >> node 1 size: 8131 MB > >> node 1 free: 6980 MB > >> node distances: > >> node 0 1 > >> 0: 10 40 > >> 1: 40 10 > >> > >> $ numastat > >> node0 node1 > >> numa_hit 108 56495 > >> numa_miss 0 0 > >> numa_foreign 0 0 > >> interleave_hit 0 4537 > >> local_node 108 31547 > >> other_node 0 24948 > >> > >> Hence, return zero explicitly for all the stats of an > >> unpopulated zone. > > > > I hope I am not just confused but I would expect that at least > > numa_foreign and other_node to be non zero. > Hmm, checking zone_statistics(): > > NUMA_FOREIGN increment uses preferred zone, which is the first in zone in > zonelist, so it will be a zone from node 1 even for allocations on cpu > associated to node 0 - assuming node 0's unpopulated zones are not included in > node 0's zonelist. But the allocation could have been requested for node 0 regardless of the amount of memory the node has. > NUMA_OTHER uses numa_node_id(), which would mean the node 0's cpus have node 1 > in their numa_node_id() ? Is that correct? numa_node_id should reflect the real node the CPU is associated with.
On 5/6/20 4:02 PM, Michal Hocko wrote: > On Wed 06-05-20 15:33:36, Vlastimil Babka wrote: >> On 5/4/20 12:26 PM, Michal Hocko wrote: >> > On Mon 04-05-20 12:33:04, Sandipan Das wrote: >> >> For unpopulated zones, the pagesets point to the common >> >> boot_pageset which can have non-zero vm_numa_stat counts. >> >> Because of this memory-less nodes end up having non-zero >> >> NUMA statistics. This can be observed on any architecture >> >> that supports memory-less NUMA nodes. >> >> >> >> E.g. >> >> >> >> $ numactl -H >> >> available: 2 nodes (0-1) >> >> node 0 cpus: 0 1 2 3 >> >> node 0 size: 0 MB >> >> node 0 free: 0 MB >> >> node 1 cpus: 4 5 6 7 >> >> node 1 size: 8131 MB >> >> node 1 free: 6980 MB >> >> node distances: >> >> node 0 1 >> >> 0: 10 40 >> >> 1: 40 10 >> >> >> >> $ numastat >> >> node0 node1 >> >> numa_hit 108 56495 >> >> numa_miss 0 0 >> >> numa_foreign 0 0 >> >> interleave_hit 0 4537 >> >> local_node 108 31547 >> >> other_node 0 24948 >> >> >> >> Hence, return zero explicitly for all the stats of an >> >> unpopulated zone. >> > >> > I hope I am not just confused but I would expect that at least >> > numa_foreign and other_node to be non zero. >> Hmm, checking zone_statistics(): >> >> NUMA_FOREIGN increment uses preferred zone, which is the first in zone in >> zonelist, so it will be a zone from node 1 even for allocations on cpu >> associated to node 0 - assuming node 0's unpopulated zones are not included in >> node 0's zonelist. > > But the allocation could have been requested for node 0 regardless of > the amount of memory the node has. Yes, if we allocate from cpu 0-3 then it should be a miss on node 0. But the zonelists are optimized in a way that they don't include empty zones - build_zonerefs_node() checks managed_zone(). As a result, node 0 zonelist has no node 0 zones, which confuses the stats code. We should probably document that numa stats are bogus on systems with memoryless nodes. This patch makes it somewhat more obvious by presenting nice zeroes on the memoryless node itself, but node 1 now include stats from node 0. >> NUMA_OTHER uses numa_node_id(), which would mean the node 0's cpus have node 1 >> in their numa_node_id() ? Is that correct? > > numa_node_id should reflect the real node the CPU is associated with. You're right, numa_node_id() is probably fine. But NUMA_OTHER is actually incremented at the zone where the allocation succeeds. This probably doesn't match Documentation/admin-guide/numastat.rst, even on a non-memoryless-node systems: other_node A process ran on this node and got memory from another node.
On Wed 06-05-20 17:09:51, Vlastimil Babka wrote: > On 5/6/20 4:02 PM, Michal Hocko wrote: > > On Wed 06-05-20 15:33:36, Vlastimil Babka wrote: > >> On 5/4/20 12:26 PM, Michal Hocko wrote: > >> > On Mon 04-05-20 12:33:04, Sandipan Das wrote: > >> >> For unpopulated zones, the pagesets point to the common > >> >> boot_pageset which can have non-zero vm_numa_stat counts. > >> >> Because of this memory-less nodes end up having non-zero > >> >> NUMA statistics. This can be observed on any architecture > >> >> that supports memory-less NUMA nodes. > >> >> > >> >> E.g. > >> >> > >> >> $ numactl -H > >> >> available: 2 nodes (0-1) > >> >> node 0 cpus: 0 1 2 3 > >> >> node 0 size: 0 MB > >> >> node 0 free: 0 MB > >> >> node 1 cpus: 4 5 6 7 > >> >> node 1 size: 8131 MB > >> >> node 1 free: 6980 MB > >> >> node distances: > >> >> node 0 1 > >> >> 0: 10 40 > >> >> 1: 40 10 > >> >> > >> >> $ numastat > >> >> node0 node1 > >> >> numa_hit 108 56495 > >> >> numa_miss 0 0 > >> >> numa_foreign 0 0 > >> >> interleave_hit 0 4537 > >> >> local_node 108 31547 > >> >> other_node 0 24948 > >> >> > >> >> Hence, return zero explicitly for all the stats of an > >> >> unpopulated zone. > >> > > >> > I hope I am not just confused but I would expect that at least > >> > numa_foreign and other_node to be non zero. > >> Hmm, checking zone_statistics(): > >> > >> NUMA_FOREIGN increment uses preferred zone, which is the first in zone in > >> zonelist, so it will be a zone from node 1 even for allocations on cpu > >> associated to node 0 - assuming node 0's unpopulated zones are not included in > >> node 0's zonelist. > > > > But the allocation could have been requested for node 0 regardless of > > the amount of memory the node has. > > Yes, if we allocate from cpu 0-3 then it should be a miss on node 0. But the > zonelists are optimized in a way that they don't include empty zones - > build_zonerefs_node() checks managed_zone(). As a result, node 0 zonelist has no > node 0 zones, which confuses the stats code. We should probably document that > numa stats are bogus on systems with memoryless nodes. This patch makes it > somewhat more obvious by presenting nice zeroes on the memoryless node itself, > but node 1 now include stats from node 0. Thanks for the clarification. So the underlying problem is that zone_statistics operates on a preferred zone rather than node. This would be fixable but I am not sure whether this is something worth bothering. Maybe it would just be more convenient to document the unfortunate memory less nodes stats situation and be done with it. Or do we have any consumers that really do care? > >> NUMA_OTHER uses numa_node_id(), which would mean the node 0's cpus have node 1 > >> in their numa_node_id() ? Is that correct? > > > > numa_node_id should reflect the real node the CPU is associated with. > > You're right, numa_node_id() is probably fine. But NUMA_OTHER is actually > incremented at the zone where the allocation succeeds. This probably doesn't > match Documentation/admin-guide/numastat.rst, even on a non-memoryless-node systems: > > other_node A process ran on this node and got memory from another node. Yeah, the documentation doesn't match the implementation. Maybe we should just fix the documentation because this has been the case for ages.
On 5/6/20 5:24 PM, Michal Hocko wrote: >> >> Yes, if we allocate from cpu 0-3 then it should be a miss on node 0. But the >> zonelists are optimized in a way that they don't include empty zones - >> build_zonerefs_node() checks managed_zone(). As a result, node 0 zonelist has no >> node 0 zones, which confuses the stats code. We should probably document that >> numa stats are bogus on systems with memoryless nodes. This patch makes it >> somewhat more obvious by presenting nice zeroes on the memoryless node itself, >> but node 1 now include stats from node 0. > > Thanks for the clarification. So the underlying problem is that zone_statistics > operates on a preferred zone rather than node. This would be fixable but > I am not sure whether this is something worth bothering. Maybe it would > just be more convenient to document the unfortunate memory less nodes > stats situation and be done with it. Or do we have any consumers that > really do care? > >> >> NUMA_OTHER uses numa_node_id(), which would mean the node 0's cpus have node 1 >> >> in their numa_node_id() ? Is that correct? >> > >> > numa_node_id should reflect the real node the CPU is associated with. >> >> You're right, numa_node_id() is probably fine. But NUMA_OTHER is actually >> incremented at the zone where the allocation succeeds. This probably doesn't >> match Documentation/admin-guide/numastat.rst, even on a non-memoryless-node systems: >> >> other_node A process ran on this node and got memory from another node. > > Yeah, the documentation doesn't match the implementation. Maybe we > should just fix the documentation because this has been the case for > ages. > How about something like this: diff --git a/Documentation/admin-guide/numastat.rst b/Documentation/admin-guide/numastat.rst index aaf1667489f8..08ec2c2bdce3 100644 --- a/Documentation/admin-guide/numastat.rst +++ b/Documentation/admin-guide/numastat.rst @@ -6,6 +6,21 @@ Numa policy hit/miss statistics All units are pages. Hugepages have separate counters. +The numa_hit, numa_miss and numa_foreign counters reflect how well processes +are able to allocate memory from nodes they prefer. If they succeed, numa_hit +is incremented on the preferred node, otherwise numa_foreign is incremented on +the preferred node and numa_miss on the node where allocation succeeded. + +Usually preferred node is the one local to the CPU where the process executes, +but restrictions such as mempolicies can change that, so there are also two +counters based on CPU local node. local_node is similar to numa_hit and is +incremented on allocation from a node by CPU on the same node. other_node is +similar to numa_miss and is incremented on the node where allocation succeeds +from a CPU from a different node. Note there is no counter analogical to +numa_foreign. + +In more detail: + =============== ============================================================ numa_hit A process wanted to allocate memory from this node, and succeeded. @@ -14,11 +29,13 @@ numa_miss A process wanted to allocate memory from another node, but ended up with memory from this node. numa_foreign A process wanted to allocate on this node, - but ended up with memory from another one. + but ended up with memory from another node. -local_node A process ran on this node and got memory from it. +local_node A process ran on this node's CPU, + and got memory from this node. -other_node A process ran on this node and got memory from another node. +other_node A process ran on a different node's CPU + and got memory from this node. interleave_hit Interleaving wanted to allocate from this node and succeeded. @@ -28,3 +45,11 @@ For easier reading you can use the numastat utility from the numactl package (http://oss.sgi.com/projects/libnuma/). Note that it only works well right now on machines with a small number of CPUs. +Note that on systems with memoryless nodes (where a node has CPUs but no +memory) the numa_hit, numa_miss and numa_foreign statistics can be skewed +heavily. In the current kernel implementation, if a process prefers a +memoryless node (i.e. because it is running on one of its local CPU), the +implementation actually treats one of the nearest nodes with memory as the +preferred node. As a result, such allocation will not increase the numa_foreign +counter on the memoryless node, and will skew the numa_hit, numa_miss and +numa_foreign statistics of the nearest node.
On Wed 06-05-20 17:50:28, Vlastimil Babka wrote: > On 5/6/20 5:24 PM, Michal Hocko wrote: > >> > >> Yes, if we allocate from cpu 0-3 then it should be a miss on node 0. But the > >> zonelists are optimized in a way that they don't include empty zones - > >> build_zonerefs_node() checks managed_zone(). As a result, node 0 zonelist has no > >> node 0 zones, which confuses the stats code. We should probably document that > >> numa stats are bogus on systems with memoryless nodes. This patch makes it > >> somewhat more obvious by presenting nice zeroes on the memoryless node itself, > >> but node 1 now include stats from node 0. > > > > Thanks for the clarification. So the underlying problem is that zone_statistics > > operates on a preferred zone rather than node. This would be fixable but > > I am not sure whether this is something worth bothering. Maybe it would > > just be more convenient to document the unfortunate memory less nodes > > stats situation and be done with it. Or do we have any consumers that > > really do care? > > > >> >> NUMA_OTHER uses numa_node_id(), which would mean the node 0's cpus have node 1 > >> >> in their numa_node_id() ? Is that correct? > >> > > >> > numa_node_id should reflect the real node the CPU is associated with. > >> > >> You're right, numa_node_id() is probably fine. But NUMA_OTHER is actually > >> incremented at the zone where the allocation succeeds. This probably doesn't > >> match Documentation/admin-guide/numastat.rst, even on a non-memoryless-node systems: > >> > >> other_node A process ran on this node and got memory from another node. > > > > Yeah, the documentation doesn't match the implementation. Maybe we > > should just fix the documentation because this has been the case for > > ages. > > > > How about something like this: > > diff --git a/Documentation/admin-guide/numastat.rst b/Documentation/admin-guide/numastat.rst > index aaf1667489f8..08ec2c2bdce3 100644 > --- a/Documentation/admin-guide/numastat.rst > +++ b/Documentation/admin-guide/numastat.rst > @@ -6,6 +6,21 @@ Numa policy hit/miss statistics > > All units are pages. Hugepages have separate counters. > > +The numa_hit, numa_miss and numa_foreign counters reflect how well processes > +are able to allocate memory from nodes they prefer. If they succeed, numa_hit > +is incremented on the preferred node, otherwise numa_foreign is incremented on > +the preferred node and numa_miss on the node where allocation succeeded. > + > +Usually preferred node is the one local to the CPU where the process executes, > +but restrictions such as mempolicies can change that, so there are also two > +counters based on CPU local node. local_node is similar to numa_hit and is > +incremented on allocation from a node by CPU on the same node. other_node is > +similar to numa_miss and is incremented on the node where allocation succeeds > +from a CPU from a different node. Note there is no counter analogical to > +numa_foreign. > + > +In more detail: > + > =============== ============================================================ > numa_hit A process wanted to allocate memory from this node, > and succeeded. > @@ -14,11 +29,13 @@ numa_miss A process wanted to allocate memory from another node, > but ended up with memory from this node. > > numa_foreign A process wanted to allocate on this node, > - but ended up with memory from another one. > + but ended up with memory from another node. > > -local_node A process ran on this node and got memory from it. > +local_node A process ran on this node's CPU, > + and got memory from this node. > > -other_node A process ran on this node and got memory from another node. > +other_node A process ran on a different node's CPU > + and got memory from this node. > > interleave_hit Interleaving wanted to allocate from this node > and succeeded. > @@ -28,3 +45,11 @@ For easier reading you can use the numastat utility from the numactl package > (http://oss.sgi.com/projects/libnuma/). Note that it only works > well right now on machines with a small number of CPUs. > > +Note that on systems with memoryless nodes (where a node has CPUs but no > +memory) the numa_hit, numa_miss and numa_foreign statistics can be skewed > +heavily. In the current kernel implementation, if a process prefers a > +memoryless node (i.e. because it is running on one of its local CPU), the > +implementation actually treats one of the nearest nodes with memory as the > +preferred node. As a result, such allocation will not increase the numa_foreign > +counter on the memoryless node, and will skew the numa_hit, numa_miss and > +numa_foreign statistics of the nearest node. This is certainly an improvement. Thanks! The question whether we can identify where bogus numbers came from would be interesting as well. Maybe those are not worth fixing but it would be great to understand them at least. I have to say that the explanation via boot_pageset is not really clear to me.
On 07/05/20 12:39 pm, Michal Hocko wrote: > On Wed 06-05-20 17:50:28, Vlastimil Babka wrote: >> [...] >> >> How about something like this: >> >> diff --git a/Documentation/admin-guide/numastat.rst b/Documentation/admin-guide/numastat.rst >> index aaf1667489f8..08ec2c2bdce3 100644 >> --- a/Documentation/admin-guide/numastat.rst >> +++ b/Documentation/admin-guide/numastat.rst >> @@ -6,6 +6,21 @@ Numa policy hit/miss statistics >> >> All units are pages. Hugepages have separate counters. >> >> +The numa_hit, numa_miss and numa_foreign counters reflect how well processes >> +are able to allocate memory from nodes they prefer. If they succeed, numa_hit >> +is incremented on the preferred node, otherwise numa_foreign is incremented on >> +the preferred node and numa_miss on the node where allocation succeeded. >> + >> +Usually preferred node is the one local to the CPU where the process executes, >> +but restrictions such as mempolicies can change that, so there are also two >> +counters based on CPU local node. local_node is similar to numa_hit and is >> +incremented on allocation from a node by CPU on the same node. other_node is >> +similar to numa_miss and is incremented on the node where allocation succeeds >> +from a CPU from a different node. Note there is no counter analogical to >> +numa_foreign. >> + >> +In more detail: >> + >> =============== ============================================================ >> numa_hit A process wanted to allocate memory from this node, >> and succeeded. >> @@ -14,11 +29,13 @@ numa_miss A process wanted to allocate memory from another node, >> but ended up with memory from this node. >> >> numa_foreign A process wanted to allocate on this node, >> - but ended up with memory from another one. >> + but ended up with memory from another node. >> >> -local_node A process ran on this node and got memory from it. >> +local_node A process ran on this node's CPU, >> + and got memory from this node. >> >> -other_node A process ran on this node and got memory from another node. >> +other_node A process ran on a different node's CPU >> + and got memory from this node. >> >> interleave_hit Interleaving wanted to allocate from this node >> and succeeded. >> @@ -28,3 +45,11 @@ For easier reading you can use the numastat utility from the numactl package >> (http://oss.sgi.com/projects/libnuma/). Note that it only works >> well right now on machines with a small number of CPUs. >> >> +Note that on systems with memoryless nodes (where a node has CPUs but no >> +memory) the numa_hit, numa_miss and numa_foreign statistics can be skewed >> +heavily. In the current kernel implementation, if a process prefers a >> +memoryless node (i.e. because it is running on one of its local CPU), the >> +implementation actually treats one of the nearest nodes with memory as the >> +preferred node. As a result, such allocation will not increase the numa_foreign >> +counter on the memoryless node, and will skew the numa_hit, numa_miss and >> +numa_foreign statistics of the nearest node. > > This is certainly an improvement. Thanks! The question whether we can > identify where bogus numbers came from would be interesting as well. > Maybe those are not worth fixing but it would be great to understand > them at least. I have to say that the explanation via boot_pageset is > not really clear to me. > The documentation update will definitely help. Thanks for that. I did collect some stack traces on a ppc64 guest for calls to zone_statistics() in case of zones that are using the boot_pageset and most of them originate from kmem_cache_init() with eventual calls to allocate_slab(). [ 0.000000] [c00000000282b690] [c000000000402d98] zone_statistics+0x138/0x1d0 [ 0.000000] [c00000000282b740] [c000000000401190] rmqueue_pcplist+0xf0/0x120 [ 0.000000] [c00000000282b7d0] [c00000000040b178] get_page_from_freelist+0x2f8/0x2100 [ 0.000000] [c00000000282bb30] [c000000000401ae0] __alloc_pages_nodemask+0x1a0/0x2d0 [ 0.000000] [c00000000282bbc0] [c00000000044b040] alloc_slab_page+0x70/0x580 [ 0.000000] [c00000000282bc20] [c00000000044b5f8] allocate_slab+0xa8/0x610 ... In the remaining cases, the sources are ftrace_init() and early_trace_init(). Unless they are useful, can we avoid incrementing stats for zones using boot_pageset inside zone_statistics()? - Sandipan
On 07/05/20 2:35 pm, Sandipan Das wrote: > > On 07/05/20 12:39 pm, Michal Hocko wrote: >> On Wed 06-05-20 17:50:28, Vlastimil Babka wrote: >>> [...] >>> >>> How about something like this: >>> >>> diff --git a/Documentation/admin-guide/numastat.rst b/Documentation/admin-guide/numastat.rst >>> index aaf1667489f8..08ec2c2bdce3 100644 >>> --- a/Documentation/admin-guide/numastat.rst >>> +++ b/Documentation/admin-guide/numastat.rst >>> @@ -6,6 +6,21 @@ Numa policy hit/miss statistics >>> >>> All units are pages. Hugepages have separate counters. >>> >>> +The numa_hit, numa_miss and numa_foreign counters reflect how well processes >>> +are able to allocate memory from nodes they prefer. If they succeed, numa_hit >>> +is incremented on the preferred node, otherwise numa_foreign is incremented on >>> +the preferred node and numa_miss on the node where allocation succeeded. >>> + >>> +Usually preferred node is the one local to the CPU where the process executes, >>> +but restrictions such as mempolicies can change that, so there are also two >>> +counters based on CPU local node. local_node is similar to numa_hit and is >>> +incremented on allocation from a node by CPU on the same node. other_node is >>> +similar to numa_miss and is incremented on the node where allocation succeeds >>> +from a CPU from a different node. Note there is no counter analogical to >>> +numa_foreign. >>> + >>> +In more detail: >>> + >>> =============== ============================================================ >>> numa_hit A process wanted to allocate memory from this node, >>> and succeeded. >>> @@ -14,11 +29,13 @@ numa_miss A process wanted to allocate memory from another node, >>> but ended up with memory from this node. >>> >>> numa_foreign A process wanted to allocate on this node, >>> - but ended up with memory from another one. >>> + but ended up with memory from another node. >>> >>> -local_node A process ran on this node and got memory from it. >>> +local_node A process ran on this node's CPU, >>> + and got memory from this node. >>> >>> -other_node A process ran on this node and got memory from another node. >>> +other_node A process ran on a different node's CPU >>> + and got memory from this node. >>> >>> interleave_hit Interleaving wanted to allocate from this node >>> and succeeded. >>> @@ -28,3 +45,11 @@ For easier reading you can use the numastat utility from the numactl package >>> (http://oss.sgi.com/projects/libnuma/). Note that it only works >>> well right now on machines with a small number of CPUs. >>> >>> +Note that on systems with memoryless nodes (where a node has CPUs but no >>> +memory) the numa_hit, numa_miss and numa_foreign statistics can be skewed >>> +heavily. In the current kernel implementation, if a process prefers a >>> +memoryless node (i.e. because it is running on one of its local CPU), the >>> +implementation actually treats one of the nearest nodes with memory as the >>> +preferred node. As a result, such allocation will not increase the numa_foreign >>> +counter on the memoryless node, and will skew the numa_hit, numa_miss and >>> +numa_foreign statistics of the nearest node. >> >> This is certainly an improvement. Thanks! The question whether we can >> identify where bogus numbers came from would be interesting as well. >> Maybe those are not worth fixing but it would be great to understand >> them at least. I have to say that the explanation via boot_pageset is >> not really clear to me. >> > > The documentation update will definitely help. Thanks for that. > I did collect some stack traces on a ppc64 guest for calls to zone_statistics() > in case of zones that are using the boot_pageset and most of them originate > from kmem_cache_init() with eventual calls to allocate_slab(). > > [ 0.000000] [c00000000282b690] [c000000000402d98] zone_statistics+0x138/0x1d0 > [ 0.000000] [c00000000282b740] [c000000000401190] rmqueue_pcplist+0xf0/0x120 > [ 0.000000] [c00000000282b7d0] [c00000000040b178] get_page_from_freelist+0x2f8/0x2100 > [ 0.000000] [c00000000282bb30] [c000000000401ae0] __alloc_pages_nodemask+0x1a0/0x2d0 > [ 0.000000] [c00000000282bbc0] [c00000000044b040] alloc_slab_page+0x70/0x580 > [ 0.000000] [c00000000282bc20] [c00000000044b5f8] allocate_slab+0xa8/0x610 > ... > > In the remaining cases, the sources are ftrace_init() and early_trace_init(). > Forgot to add that this happens during the period between zone_pcp_init() and setup_zone_pageset(). - Sandipan
On 5/7/20 11:08 AM, Sandipan Das wrote: >>> >>> This is certainly an improvement. Thanks! The question whether we can >>> identify where bogus numbers came from would be interesting as well. >>> Maybe those are not worth fixing but it would be great to understand >>> them at least. I have to say that the explanation via boot_pageset is >>> not really clear to me. >>> >> >> The documentation update will definitely help. Thanks for that. Thanks both, will send a proper patch. >> I did collect some stack traces on a ppc64 guest for calls to zone_statistics() >> in case of zones that are using the boot_pageset and most of them originate >> from kmem_cache_init() with eventual calls to allocate_slab(). >> >> [ 0.000000] [c00000000282b690] [c000000000402d98] zone_statistics+0x138/0x1d0 >> [ 0.000000] [c00000000282b740] [c000000000401190] rmqueue_pcplist+0xf0/0x120 >> [ 0.000000] [c00000000282b7d0] [c00000000040b178] get_page_from_freelist+0x2f8/0x2100 >> [ 0.000000] [c00000000282bb30] [c000000000401ae0] __alloc_pages_nodemask+0x1a0/0x2d0 >> [ 0.000000] [c00000000282bbc0] [c00000000044b040] alloc_slab_page+0x70/0x580 >> [ 0.000000] [c00000000282bc20] [c00000000044b5f8] allocate_slab+0xa8/0x610 >> ... >> >> In the remaining cases, the sources are ftrace_init() and early_trace_init(). >> > > Forgot to add that this happens during the period between zone_pcp_init() and setup_zone_pageset(). If you can identify a good moment during the boot when the boot_pageset stops being used, maybe we could just reset its stats to zero once at that point and be done with it. > - Sandipan >
On 07/05/20 4:37 pm, Vlastimil Babka wrote: > On 5/7/20 11:08 AM, Sandipan Das wrote: >>>> >>>> This is certainly an improvement. Thanks! The question whether we can >>>> identify where bogus numbers came from would be interesting as well. >>>> Maybe those are not worth fixing but it would be great to understand >>>> them at least. I have to say that the explanation via boot_pageset is >>>> not really clear to me. >>>> >>> >>> The documentation update will definitely help. Thanks for that. > > Thanks both, will send a proper patch. > >>> I did collect some stack traces on a ppc64 guest for calls to zone_statistics() >>> in case of zones that are using the boot_pageset and most of them originate >>> from kmem_cache_init() with eventual calls to allocate_slab(). >>> >>> [ 0.000000] [c00000000282b690] [c000000000402d98] zone_statistics+0x138/0x1d0 >>> [ 0.000000] [c00000000282b740] [c000000000401190] rmqueue_pcplist+0xf0/0x120 >>> [ 0.000000] [c00000000282b7d0] [c00000000040b178] get_page_from_freelist+0x2f8/0x2100 >>> [ 0.000000] [c00000000282bb30] [c000000000401ae0] __alloc_pages_nodemask+0x1a0/0x2d0 >>> [ 0.000000] [c00000000282bbc0] [c00000000044b040] alloc_slab_page+0x70/0x580 >>> [ 0.000000] [c00000000282bc20] [c00000000044b5f8] allocate_slab+0xa8/0x610 >>> ... >>> >>> In the remaining cases, the sources are ftrace_init() and early_trace_init(). >>> >> >> Forgot to add that this happens during the period between zone_pcp_init() and setup_zone_pageset(). > > If you can identify a good moment during the boot when the boot_pageset stops > being used, maybe we could just reset its stats to zero once at that point and > be done with it. > I think that would be in setup_per_cpu_pageset(). After the pagesets for the populated zone have been allocated. - Sandipan
diff --git a/include/linux/vmstat.h b/include/linux/vmstat.h index 292485f3d24d..55a68b379a2c 100644 --- a/include/linux/vmstat.h +++ b/include/linux/vmstat.h @@ -159,6 +159,21 @@ static inline unsigned long zone_numa_state_snapshot(struct zone *zone, long x = atomic_long_read(&zone->vm_numa_stat[item]); int cpu; + /* + * Initially, the pageset of all zones are set to point to the + * boot_pageset. The real pagesets are allocated later but only + * for the populated zones. Unpopulated zones still continue + * using the boot_pageset. + * + * Before the real pagesets are allocated, the boot_pageset's + * vm_numa_stat counters can get incremented. This affects the + * unpopulated zones which end up with non-zero stats despite + * having no memory associated with them. For such cases, + * return zero explicitly. + */ + if (!populated_zone(zone)) + return 0; + for_each_online_cpu(cpu) x += per_cpu_ptr(zone->pageset, cpu)->vm_numa_stat_diff[item];
For unpopulated zones, the pagesets point to the common boot_pageset which can have non-zero vm_numa_stat counts. Because of this memory-less nodes end up having non-zero NUMA statistics. This can be observed on any architecture that supports memory-less NUMA nodes. E.g. $ numactl -H available: 2 nodes (0-1) node 0 cpus: 0 1 2 3 node 0 size: 0 MB node 0 free: 0 MB node 1 cpus: 4 5 6 7 node 1 size: 8131 MB node 1 free: 6980 MB node distances: node 0 1 0: 10 40 1: 40 10 $ numastat node0 node1 numa_hit 108 56495 numa_miss 0 0 numa_foreign 0 0 interleave_hit 0 4537 local_node 108 31547 other_node 0 24948 Hence, return zero explicitly for all the stats of an unpopulated zone. Signed-off-by: Sandipan Das <sandipan@linux.ibm.com> --- include/linux/vmstat.h | 15 +++++++++++++++ 1 file changed, 15 insertions(+)