Date:   Mon, 27 Mar 2023 10:11:19 -0400
From:   Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
To:     Aaron Lu <aaron.lu@...el.com>
Cc:     Peter Zijlstra <peterz@...radead.org>, linux-kernel@...r.kernel.org
Subject: Re: rq lock contention due to commit af7f588d8f73

On 2023-03-27 10:04, Aaron Lu wrote:
> On Mon, Mar 27, 2023 at 09:20:44AM -0400, Mathieu Desnoyers wrote:
>> On 2023-03-27 04:05, Aaron Lu wrote:
>>> Hi Mathieu,
>>>
>>> I was doing some optimization work[1] for kernel scheduler using a
>>> database workload: sysbench+postgres and before I submit my work, I
>>> rebased my patch on top of latest v6.3-rc kernels to see if everything
>>> still works expected and then I found rq's lock became very heavily
>>> contended as compared to v6.2 based kernels.
>>>
>>> Using the above mentioned workload, before commit af7f588d8f73("sched:
>>> Introduce per-memory-map concurrency ID"), the profile looked like:
>>>
>>>        7.30%     0.71%  [kernel.vmlinux]            [k] __schedule
>>>        0.03%     0.03%  [kernel.vmlinux]            [k] native_queued_spin_lock_slowpath
>>>
>>> After that commit:
>>>
>>>       49.01%     0.87%  [kernel.vmlinux]            [k] __schedule
>>>       43.20%    43.18%  [kernel.vmlinux]            [k] native_queued_spin_lock_slowpath
>>>
>>> The above profile was captured with sysbench's nr_threads set to 56; if
>>> I used more thread number, the contention would be more severe on that
>>> 2sockets/112core/224cpu Intel Sapphire Rapids server.
>>>
>>> The docker image I used to do optimization work is not available outside
>>> but I managed to reproduce this problem using only publicaly available
>>> stuffs, here it goes:
>>> 1 docker pull postgres
>>> 2 sudo docker run --rm --name postgres-instance -e POSTGRES_PASSWORD=mypass -e POSTGRES_USER=sbtest -d postgres -c shared_buffers=80MB -c max_connections=250
>>> 3 go inside the container
>>>     sudo docker exec -it $the_just_started_container_id bash
>>> 4 install sysbench inside container
>>>     sudo apt update and sudo apt install sysbench
>>> 5 prepare
>>>     root@...tainer:/# sysbench --db-driver=pgsql --pgsql-user=sbtest --pgsql_password=mypass --pgsql-db=sbtest --pgsql-port=5432 --tables=16 --table-size=10000 --threads=56 --time=60 --report-interval=2 /usr/share/sysbench/oltp_read_only.lua prepare
>>> 6 run
>>>     root@...tainer:/# sysbench --db-driver=pgsql --pgsql-user=sbtest --pgsql_password=mypass --pgsql-db=sbtest --pgsql-port=5432 --tables=16 --table-size=10000 --threads=56 --time=60 --report-interval=2 /usr/share/sysbench/oltp_read_only.lua run
>>>
>>> Let it warm up a little bit and after 10-20s you can do profile and see
>>> the increased rq lock contention. You may need a machine that has at
>>> least 56 cpus to see this, I didn't try on other machines.
>>>
>>> Feel free to let me know if you need any other info.
>>
>> While I setup my dev machine with this reproducer, here are a few
>> questions to help figure out the context:
>>
>> I understand that pgsql is a multi-process database. Is it strictly
>> single-threaded per-process, or does each process have more than
>> one thread ?
> 
> I do not know the details of Postgres, according to this:
> https://wiki.postgresql.org/wiki/FAQ#How_does_PostgreSQL_use_CPU_resources.3F
> I think it is single-threaded per-process.
> 
> The client, sysbench, is single process multi-threaded IIUC.
> 
>>
>> I understand that your workload is scheduling between threads which
>> belong to different processes. Are there more heavily active threads
>> than there are scheduler runqueues (CPUs) on your machine ?
> 
> In the reproducer I described above, 56 threads are started on the
> client side and if each client thread is served by a server process,
> there would be about 112 tasks. I don't think the client thread and
> the server process are active at the same time but even if they are,
> 112 is still smaller than the machine's CPU number: 224.
> 
>>
>> When I developed the mm_cid feature, I originally implemented two additional
>> optimizations:
>>
>>      Additional optimizations can be done if the spin locks added when
>>      context switching between threads belonging to different memory maps end
>>      up being a performance bottleneck. Those are left out of this patch
>>      though. A performance impact would have to be clearly demonstrated to
>>      justify the added complexity.
>>
>> I suspect that your workload demonstrates the need for at least one of those
>> optimizations. I just wonder if we are in a purely single-threaded scenario
>> for each process, or if each process has many threads.
> 
> My understanding is: the server side is single threaded and the client
> side is multi threaded.

Indeed, I just validated this by successfully running your reproducer 
locally, and htop confirms that the client is single-process with many 
threads, and the server is multi-process, each with a single thread.

So in this case, the simple "single-threaded process" optimization would 
not work, because the client-side is multi-threaded. So the scheduler 
will switch back and forth between the client process and the server 
processes.

So this appears to call for my mm_cid runqueue cache.

Thanks,

Mathieu


> 
> Thanks,
> Aaron

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

Powered by blists - more mailing lists