[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <1a1e06d4-7f41-4f37-a9b5-d1610e8d8669@efficios.com>
Date: Wed, 4 Sep 2024 14:28:05 -0400
From: Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
To: Yury Norov <yury.norov@...il.com>
Cc: Peter Zijlstra <peterz@...radead.org>, Ingo Molnar <mingo@...hat.com>,
linux-kernel@...r.kernel.org, Valentin Schneider <vschneid@...hat.com>,
Mel Gorman <mgorman@...e.de>, Steven Rostedt <rostedt@...dmis.org>,
Vincent Guittot <vincent.guittot@...aro.org>,
Dietmar Eggemann <dietmar.eggemann@....com>, Ben Segall
<bsegall@...gle.com>, Rasmus Villemoes <linux@...musvillemoes.dk>,
Dmitry Vyukov <dvyukov@...gle.com>, Marco Elver <elver@...gle.com>
Subject: Re: [RFC PATCH 2/2] sched: Improve cache locality of RSEQ concurrency
IDs for intermittent workloads
On 2024-09-04 11:50, Mathieu Desnoyers wrote:
> On 2024-09-04 11:24, Yury Norov wrote:
[...]
>>
>> This all doesn't look like a hot path. And anyways, speculating around
>> performance without numbers on hands sounds cheap.
>
> This is done whenever userspace invokes sched_setaffinity, or changes
> its cgroup cpuset. It may not be the most important fast-path in the
> world, but I expect some workloads to issue sched_setaffinity whenever
> they create a thread, so it's not a purely slow-path either.
>
>> In my experience, iterators with a very lightweight payload are ~100
>> times slower comparing to dedicated bitmap ops. Check this for example:
>> 3cea8d4753277.
>>
>> If you're really cared about performance here, I'd suggest you to
>> compare your iterators approach with something like this:
>>
>> cpumask_or(mm_allowed, mm_allowed, cpumask);
>> atomic_set(&mm->nr_cpus_allowed, cpumask_weight(mm_allowed);
Here are the benchmark results. Each test use two entirely filled
bitmaps as input to mimic the common scenario for cpus allowed
being updated with a subset of the original process CPUs allowed,
and also the common case where the initial cpumask is filled.
#define BITMAP_LEN (4096UL * 8 * 10)
(len = BITMAP_LEN)
* Approach 1:
int nr_set = 0;
for_each_andnot_bit(bit, bitmap, bitmap2, len)
nr_set += !test_and_set_bit(bit, bitmap2);
if (nr_set)
atomic_add(nr_set, &total);
Time: 4680 ns
* Approach 2:
int nr_set = 0;
for_each_set_bit(bit, bitmap, len)
nr_set += !test_and_set_bit(bit, bitmap2);
if (nr_set)
atomic_add(nr_set, &total);
Time: 1791537 ns
* Approach 3:
mutex_lock(&lock);
bitmap_or(bitmap2, bitmap, bitmap2, len);
atomic_set(&total, bitmap_weight(bitmap2, len));
mutex_unlock(&lock);
Time: 79591 ns
So approach 1 is 382 times faster than approach 2, and 17 times
faster than approach 3. And this is only single-threaded,
I expect approaches 2&3 to perform even worse when contended
due to the many fully ordered test_and_set_bit (2) and due to
locking (3).
The test hardware is a AMD EPYC 9654 96-Core Processor.
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
Powered by blists - more mailing lists