Received-SPF: softfail (google.com: domain of transitioning linux-kernel-owner@vger.kernel.org does not designate 23.128.96.19 as permitted sender) client-ip=23.128.96.19;
Message-ID: <bd95a0f0-5589-2d9e-8fb0-a66322e556e4@scylladb.com>
Date:   Wed, 30 Mar 2022 14:01:21 +0300
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.4.0
Content-Language: en-US
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     Asias He <asias@scylladb.com>, linux-kernel@vger.kernel.org
From:   Avi Kivity <avi@scylladb.com>
Subject: sched_min_granuality_ns exile into debugfs
Organization: ScyllaDB
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi Peter,


In 8a99b683 ("sched: Move SCHED_DEBUG sysctl to debugfs"), you moved

sched_min_granularity_ns to debugfs, citing that it is debug-only (true)

and undocumented (it is documented in sched-design-CFS.rst, under

the old name).


This breaks my application, Scylla[1]. We use sched_min_granularity_ns

to reduce the chances that a high networking backlog will starve the

application thread. It is a thread-per-core design, so we won't find another

core for the application, they are all busy (and besides, the application

threads are pinned).


In addition to sched_min_granularity_ns, we also tune a few other

sysctls:


# Prevent auto-scaling from doing anything to our tunables
kernel.sched_tunable_scaling = 0

# Preempt sooner
kernel.sched_min_granularity_ns = 500000

# Don't delay unrelated workloads
kernel.sched_wakeup_granularity_ns = 450000

# Schedule all tasks in this period
kernel.sched_latency_ns = 1000000

# autogroup seems to prevent sched_latency_ns from being respected
kernel.sched_autogroup_enabled = 0

# Disable numa balancing
kernel.numa_balancing = 0


While we can adapt to the move, I would much prefer it if the old location

was restored. I think it even makes sense to make this a non-debug tunable;

it helps to application to be more responsive without using the realtime

class, which is its own can of worms (and will likely result in reduced 
throughput).


[1] https://github.com/scylladb/scylla