linux-kernel - Re: [PATCH 0/9] rcu: Support for lazy callbacks on !CONFIG_RCU_NOCB

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <ZKVk4jUtsPdfgjjJ@localhost.localdomain>
Date:   Wed, 5 Jul 2023 14:41:06 +0200
From:   Frederic Weisbecker <frederic@...nel.org>
To:     "Paul E . McKenney" <paulmck@...nel.org>
Cc:     LKML <linux-kernel@...r.kernel.org>, rcu <rcu@...r.kernel.org>,
        Uladzislau Rezki <urezki@...il.com>,
        Neeraj Upadhyay <quic_neeraju@...cinc.com>,
        Joel Fernandes <joel@...lfernandes.org>,
        Giovanni Gherdovich <ggherdovich@...e.cz>
Subject: Re: [PATCH 0/9] rcu: Support for lazy callbacks on
 !CONFIG_RCU_NOCB_CPU

Le Wed, May 31, 2023 at 12:17:27PM +0200, Frederic Weisbecker a écrit :
> Hi,
> 
> Here is a first attempt at providing support for lazy callbacks on
> non-offloaded CPUs. I have measured interesting results on my machine
> when it is mostly idle (just an idle ssh connection kept alive) with
> a gain of +10% idle time. This needs more serious testing though,
> especially to make sure that no performance regression is introduced.

So after refining my measurements, it actually doesn't seem to bring
much.

Here's what I measured (it's an average after 100 runs):

* 0.01% more idle time during a "sleep 10"
* introduce 3% performance regression during an "scp" on the tools directory
  of the kernel source.
* Brings less than 1% performance improvement with a "make" on tools/perf

It only brings unconditional improvement when rcu boost is enabled (callbacks
offloaded to rcuc) but still below 1%.

The reason for the performance penalty is unclear, I initially thought it
was due to the batch execution delaying other softirqs vectors but it doesn't
look that way after all.

So for now I'm only going to re-iterate the cleanups in the beginning
of the patchset and I'll wait for a good reason before ever proceeding with the
rest.

Thanks.