Re: contention on pwq->pool->lock under heavy NFS workload

[Chuck Lever III <chuck.lever@oracle.com>]

> On Jun 21, 2023, at 5:28 PM, Tejun Heo <tj@kernel.org> wrote:
> 
> Hello,
> 
> On Wed, Jun 21, 2023 at 03:26:22PM +0000, Chuck Lever III wrote:
>> lock_stat reports that the pool->lock kernel/workqueue.c:1483 is the highest
>> contended lock on my test NFS client. The issue appears to be that the three
>> NFS-related workqueues, rpciod_workqueue, xprtiod_workqueue, and nfsiod all
>> get placed in the same worker_pool, so they have to fight over one pool lock.
>> 
>> I notice that ib_comp_wq is allocated with the same flags, but I don't see
>> significant contention there, and a trace_printk in __queue_work shows that
>> work items queued on that WQ seem to alternate between at least two different
>> worker_pools.
>> 
>> Is there a preferred way to ensure the NFS WQs get spread a little more fairly
>> amongst the worker_pools?
> 
> Can you share the output of lstopo on the test machine?
> 
> The following branch has pending workqueue changes which makes unbound
> workqueues finer grained by default and a lot more flexible in how they're
> segmented.
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git affinity-scopes-v2
> 
> Can you please test with the brnach? If the default doesn't improve the
> situation, you can set WQ_SYSFS on the affected workqueues and change their
> scoping by writing to /sys/devices/virtual/WQ_NAME/affinity_scope. Please
> take a look at
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git/tree/Documentation/core-api/workqueue.rst?h=affinity-scopes-v2#n350
> 
> for more details.

The good news:

On stock 6.4-rc7:

fio 8k [r=108k,w=46.9k IOPS]

On the affinity-scopes-v2 branch (with no other tuning):

fio 8k [r=130k,w=55.9k IOPS]


The bad news:

pool->lock is still the hottest lock on the system during the test.


I'll try some of the alternate scope settings this afternoon.


--
Chuck Lever