On Thu, May 29, 2014 at 02:56:15PM -0500, Christoph Lameter wrote:
>
> V6->V7
> - Remove /sysfs support and avoid the large cpumask definition.
>
> V5->V6:
> - Shepherd thread as a general worker thread. This means
> that the general mechanism to control worker thread
> cpu use by Frederic Weisbecker is necessary to
> restrict the shepherd thread to the cpus not used
> for low latency tasks. Hopefully that is ready to be
> merged soon. No need anymore to have a specific
> cpu be the housekeeper cpu.
>
> V4->V5:
> - Shepherd thread on a specific cpu (HOUSEKEEPING_CPU).
> - Incorporate Andrew's feedback
> - Work out the races.
> - Make visible which CPUs have stat updates switched off
> in /sys/devices/system/cpu/stat_off
>
> V3->V4:
> - Make the shepherd task not deferrable. It runs on the tick cpu
> anyways. Deferral could get deltas too far out of sync if
> vmstat operations are disabled for a certain processor.
>
> V2->V3:
> - Introduce a new tick_get_housekeeping_cpu() function. Not sure
> if that is exactly what we want but it is a start. Thomas?
> - Migrate the shepherd task if the output of
> tick_get_housekeeping_cpu() changes.
> - Fixes recommended by Andrew.
>
> V1->V2:
> - Optimize the need_update check by using memchr_inv.
> - Clean up.
>
> vmstat workers are used for folding counter differentials into the
> zone, per node and global counters at certain time intervals.
> They currently run at defined intervals on all processors which will
> cause some holdoff for processors that need minimal intrusion by the
> OS.
>
> The current vmstat_update mechanism depends on a deferrable timer
> firing every other second by default which registers a work queue item
> that runs on the local CPU, with the result that we have 1 interrupt
> and one additional schedulable task on each CPU every 2 seconds
> If a workload indeed causes VM activity or multiple tasks are running
> on a CPU, then there are probably bigger issues to deal with.
>
> However, some workloads dedicate a CPU for a single CPU bound task.
> This is done in high performance computing, in high frequency
> financial applications, in networking (Intel DPDK, EZchip NPS) and with
> the advent of systems with more and more CPUs over time, this may become
> more and more common to do since when one has enough CPUs
> one cares less about efficiently sharing a CPU with other tasks and
> more about efficiently monopolizing a CPU per task.
>
> The difference of having this timer firing and workqueue kernel thread
> scheduled per second can be enormous. An artificial test measuring the
> worst case time to do a simple "i++" in an endless loop on a bare metal
> system and under Linux on an isolated CPU with dynticks and with and
> without this patch, have Linux match the bare metal performance
> (~700 cycles) with this patch and loose by couple of orders of magnitude
> (~200k cycles) without it[*]. The loss occurs for something that just
> calculates statistics. For networking applications, for example, this
> could be the difference between dropping packets or sustaining line rate.
>
> Statistics are important and useful, but it would be great if there
> would be a way to not cause statistics gathering produce a huge
> performance difference. This patche does just that.
>
> This patch creates a vmstat shepherd worker that monitors the
> per cpu differentials on all processors. If there are differentials
> on a processor then a vmstat worker local to the processors
> with the differentials is created. That worker will then start
> folding the diffs in regular intervals. Should the worker
> find that there is no work to be done then it will make the shepherd
> worker monitor the differentials again.
>
> With this patch it is possible then to have periods longer than
> 2 seconds without any OS event on a "cpu" (hardware thread).
>
> Reviewed-by: Gilad Ben-Yossef <[email protected]>
> Signed-off-by: Christoph Lameter <[email protected]>
So after the cpumask_var_t conversion I have no other concern except
perhaps that the scan may bring some overhead on workloads that don't
care about isolation. You might want to make it optional. But I let you
check that.
And I can't judge much the -mm internal changes. But other than that, it looks good to me.
Thanks.
On Tue, 3 Jun 2014, Frederic Weisbecker wrote:
> So after the cpumask_var_t conversion I have no other concern except
> perhaps that the scan may bring some overhead on workloads that don't
> care about isolation. You might want to make it optional. But I let you
> check that.
Testing so far indicates that typical loads have spurts of kernel usage
which need vmstat but otherwise there are large segments of processing
that do not need the vmstat worker. It seems that this change is generally
helpful.
On Tue, 3 Jun 2014, Frederic Weisbecker wrote:
> So after the cpumask_var_t conversion I have no other concern except
Is there some way to observe which worker threads are queued on which
processor? I see nothing in /sys/devices/virtual/workqueues (urg should be
/sys/kernel/workqueues) that shows that?
On Tue, Jun 03, 2014 at 11:38:02AM -0500, Christoph Lameter wrote:
> On Tue, 3 Jun 2014, Frederic Weisbecker wrote:
>
> > So after the cpumask_var_t conversion I have no other concern except
>
> Is there some way to observe which worker threads are queued on which
> processor? I see nothing in /sys/devices/virtual/workqueues (urg should be
> /sys/kernel/workqueues) that shows that?
Yeah you can see that with workqueue tracing events.
Check out /sys/kernel/debug/tracing/events/workqueue/ and more specifically
workqueue_execute_start/ and workqueue_queue_work/.