Mel Gorman <mgorman <at> suse.de> writes:
Hello, Mel.
> This continues to build on the previous feedback and further testing and
> I'm hoping this can be finalised relatively soon. False sharing is still
> a major problem but I still think it deserves its own series. Minimally I
> think the fact that we are now scanning shared pages without much additional
> system overhead is a big step in the right direction.
Unfortunately, I am not seeing any improvement for situations where there
are more than one multi-threaded processes running. These would be one or
more processes per node, like a multi-JVM JBB test. Or a
multi-virtual-machine test (in my case running dbench).
One of the concerns I have is that the way numa_preferred_node decision is
made. What I am seeing is that when I run workloads as described above,
which threads of same process are likely to have accesses to same memory
areas, those threads are not grouped together and given a common preferred
node. What happens is that the threads get different numa_preferred_nid
values, and they continue to access the same pool of memory, and just bounce
the memory back and forth.
I added a few stats to show the problem:
faults(K) is the total fault count and not decayed
aged_faults are the decayed, per node faults
switches is the number of times the preferred_nid changed
The data below is for 1 of 4 active Qemu(KVM VM) with 20 vCPUs each. The
host is a 4 socket, 40 core, 80 thread westmere-EX. Ideally each of the VMs
would be run in just one NUMA node. This data was collected after dbench
was running for 5 minutes). Notice that the tasks from a single VM are
spread across all nodes.
pid: 57050
tid pref_nid switches faults(K) aged_faults(K) full_scans
57050 2 2 3.5 .1,.2,.2,0 4
57053 0 1 85.0 5.6,2.0,4.5,.8 4
57054 2 2 60.5 .7,.2,2.8,2.0 4
57055 0 2 71.0 5.0,1.0,2.0,.1 4
57056 2 2 68.5 2.1,.3,2.5,.7 4
57057 1 2 78.5 0,2.0,.3,1.1 4
57058 1 2 72.0 .3,2.0,1.5,.5 4
57059 3 2 52.0 0,.5,1.2,1.5 4
57060 1 1 42.0 .2,2.8,0,0 4
57061 2 2 58.0 .7,0,1.8,.8 4
57062 0 2 48.0 2.4,1.0,.2,.6 4
57063 2 1 45.0 .5,.2,2.7,0 4
57064 3 2 73.5 1.5,.5,.3,2.3 4
57065 0 1 70.0 4.2,.5,0,1.7 4
57066 3 2 65.0 0,0,1.5,3.2 4
57067 2 2 52.9 .2,0,5.0,4.2 4
57068 1 2 50.0 0,1.5,0,1.5 4
57069 1 2 45.5 .5,2.2,.8,0 4
57070 1 2 54.0 .1,2.0,1.0,0 4
57071 0 2 53.0 3.5,0,1.0,1.1 4
57072 0 2 45.6 2.5,0,0,0 4
57074 -1 0 0.0 0 0
69199 2 1 0.0 0,0,0,0 4
69200 0 3 0.0 0,0,0,0 4
69201 0 3 0.0 0,0,0,0 4
Here's one more Qemu process which has pretty much the same issue (all 4
look the same)
pid: 57108
tid pref_nid switches faults(K) aged_faults(K) full_scans
57108 3 1 2.3 0,.1,0,.3 5
57111 3 1 81.2 3.7,4.5,0,13.2 5
57112 3 2 55.0 0,1.5,.3,8.3 5
57113 1 2 108.51.7,27.1,1.5,1.0 5
57114 2 3 45.5 2.5,.3,4.2,1.8 5
57115 2 3 41.5 .2,.7,3.2,.2 5
57116 1 3 58.0 1.0,5.7,.1,4.7 5
57117 0 4 44.3 3.5,1.0,.7,.8 5
57118 2 3 47.5 0,0,6.7,.5 5
57119 1 3 52.5 1.7,5.6,.1,.5 5
57120 0 3 51.5 7.0,.5,.1,2.0 5
57121 3 2 38.0 1.0,0,0,4.7 5
57122 3 3 38.0 .1,1.5,0,6.5 5
57123 1 2 43.5 0,4.5,1.1,.1 5
57124 1 3 40.0 0,5.5,0,.2 5
57125 1 3 30.0 0,3.5,.7,0 5
57126 3 2 46.4 0,.1,.1,6.5 5
57127 1 4 33.5 0,2.0,1.7,1.1 5
57128 2 3 33.5 0,.2,5.2,1.0 5
57129 2 1 42.5 0,0,3.2,0 5
57130 0 2 26.5 2.1,0,1.3,0 5
57132 -1 0 0.0 0 0
69202 -1 0 0.0 0,0,0,0 5
69203 3 1 0.0 0,0,0,0 5
I think we need to establish some sort of "attracting" and "opposing" forces
when deciding to choose a numa_perferred_nid. Even if tasks of a same
process are not necessarily sharing memory accesses at the same time, it
might be beneficial to give them more "closeness" than say, a task from a
different process which has zero chance of having common accesses (no actual
shared memory between them).
One more thing regarding the common preferred_nid: this does not have to
restrict us to just one preferred_nid for each process. We certainly have
some situations where a multi-threaded process is so big that it spans the
entire system (single-JVM JBB). In that case we can try to identify
multiple thread groups within that process, and each can have a different
preferred_nid. Might even discover that the number of thread-groups =
number of threads (complete isolation in memory access pattern between the
threads).
Here's the stats patch (I am sending this via gmane, so this patch might get
garbled):
signed-off-by: Andrew Theurer <habanero@xxxxxxxxxxxxxxxxxx>
diff -Naurp linux-git.numabalv5-host/fs/proc/array.c
linux-git.numabalv5-stats-host/fs/proc/array.c
--- linux-git.numabalv5-host/fs/proc/array.c 2013-07-30 08:50:12.039076879 -0500
+++ linux-git.numabalv5-stats-host/fs/proc/array.c 2013-07-30
08:40:32.539036816 -0500
@@ -344,6 +344,26 @@ static inline void task_seccomp(struct s
#endif
}
+static inline void task_numa(struct seq_file *m,
+ struct task_struct *p)
+{
+ int nid;
+
+ if (p->raw_numa_faults) {
+ seq_printf(m,"full_scans:\t%d\n", ACCESS_ONCE(p->mm->numa_scan_seq));
+ seq_printf(m,"raw_faults:\t%lu\n", p->raw_numa_faults);
+ }
+ if (p->numa_faults) {
+ seq_printf(m,"aged_faults:\t");
+ for (nid = 0; nid < nr_node_ids; nid++)
+ seq_printf(m, "%lu ", p->numa_faults[task_faults_idx(nid,
1)]);
+ }
+ seq_printf(m, "\npreferred_nid:\t%d\n"
+ "preferred_nid_switches:\t%d\n",
+ p->numa_preferred_nid,
+ p->numa_preferred_nid_switches);
+}
+
static inline void task_context_switch_counts(struct seq_file *m,
struct task_struct *p)
{
@@ -363,6 +383,16 @@ static void task_cpus_allowed(struct seq
seq_putc(m, '\n');
}
+int proc_pid_numa(struct seq_file *m, struct pid_namespace *ns,
+ struct pid *pid, struct task_struct *task)
+{
+ struct mm_struct *mm = get_task_mm(task);
+
+ if (mm)
+ task_numa(m, task);
+ return 0;
+}
+
int proc_pid_status(struct seq_file *m, struct pid_namespace *ns,
struct pid *pid, struct task_struct *task)
{
diff -Naurp linux-git.numabalv5-host/fs/proc/base.c
linux-git.numabalv5-stats-host/fs/proc/base.c
--- linux-git.numabalv5-host/fs/proc/base.c 2013-07-26 12:51:58.103854539 -0500
+++ linux-git.numabalv5-stats-host/fs/proc/base.c 2013-07-25
08:39:20.424140726 -0500
@@ -2853,6 +2853,7 @@ static const struct pid_entry tid_base_s
REG("environ", S_IRUSR, proc_environ_operations),
INF("auxv", S_IRUSR, proc_pid_auxv),
ONE("status", S_IRUGO, proc_pid_status),
+ ONE("numa", S_IRUGO, proc_pid_numa),
ONE("personality", S_IRUGO, proc_pid_personality),
INF("limits", S_IRUGO, proc_pid_limits),
#ifdef CONFIG_SCHED_DEBUG
diff -Naurp linux-git.numabalv5-host/include/linux/sched.h
linux-git.numabalv5-stats-host/include/linux/sched.h
--- linux-git.numabalv5-host/include/linux/sched.h 2013-07-26
12:54:26.141858634 -0500
+++ linux-git.numabalv5-stats-host/include/linux/sched.h 2013-07-25
09:09:09.571245867 -0500
@@ -1522,8 +1522,10 @@ struct task_struct {
* decay and these values are copied.
*/
unsigned long *numa_faults_buffer;
+ unsigned long raw_numa_faults;
int numa_preferred_nid;
+ int numa_preferred_nid_switches;
#endif /* CONFIG_NUMA_BALANCING */
struct rcu_head rcu;
@@ -1604,6 +1606,10 @@ struct task_struct {
#ifdef CONFIG_NUMA_BALANCING
extern void task_numa_fault(int last_node, int node, int pages, bool migrated);
extern void set_numabalancing_state(bool enabled);
+static inline int task_faults_idx(int nid, int priv)
+{
+ return 2 * nid + priv;
+}
#else
static inline void task_numa_fault(int last_node, int node, int pages,
diff -Naurp linux-git.numabalv5-host/kernel/sched/core.c
linux-git.numabalv5-stats-host/kernel/sched/core.c
--- linux-git.numabalv5-host/kernel/sched/core.c 2013-07-26
12:53:17.232856728 -0500
+++ linux-git.numabalv5-stats-host/kernel/sched/core.c 2013-07-26
12:57:24.452863565 -0500
@@ -1593,9 +1593,11 @@ static void __sched_fork(struct task_str
p->numa_migrate_seq = 0;
p->numa_scan_period = sysctl_numa_balancing_scan_delay;
p->numa_preferred_nid = -1;
+ p->numa_preferred_nid_switches = 0;
p->numa_work.next = &p->numa_work;
p->numa_faults = NULL;
p->numa_faults_buffer = NULL;
+ p->raw_numa_faults = 0;
#endif /* CONFIG_NUMA_BALANCING */
}
diff -Naurp linux-git.numabalv5-host/kernel/sched/fair.c
linux-git.numabalv5-stats-host/kernel/sched/fair.c
--- linux-git.numabalv5-host/kernel/sched/fair.c 2013-07-26
12:54:12.717858262 -0500
+++ linux-git.numabalv5-stats-host/kernel/sched/fair.c 2013-07-26
12:56:32.804862137 -0500
@@ -841,11 +841,6 @@ static unsigned int task_scan_max(struct
*/
unsigned int sysctl_numa_balancing_settle_count __read_mostly = 3;
-static inline int task_faults_idx(int nid, int priv)
-{
- return 2 * nid + priv;
-}
-
static unsigned long source_load(int cpu, int type);
static unsigned long target_load(int cpu, int type);
static unsigned long power_of(int cpu);
@@ -1026,6 +1021,7 @@ static void task_numa_placement(struct t
/* Queue task on preferred node if possible */
p->numa_preferred_nid = max_nid;
+ p->numa_preferred_nid_switches++;
p->numa_migrate_seq = 0;
numa_migrate_preferred(p);
@@ -1094,6 +1090,8 @@ void task_numa_fault(int last_nidpid, in
/* Record the fault, double the weight if pages were migrated */
p->numa_faults_buffer[task_faults_idx(node, priv)] += pages << migrated;
+
+ p->raw_numa_faults += pages;
}
static void reset_ptenuma_scan(struct task_struct *p)
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>