Date: Mon, 15 Oct 2018 12:19:24 -0700 From: Gleb Smirnoff <glebius@freebsd.org> To: Mateusz Guzik <mjg@freebsd.org> Cc: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: Re: svn commit: r339349 - in head/sys/amd64: amd64 include Message-ID: <20181015191924.GF1044@FreeBSD.org> In-Reply-To: <201810132118.w9DLIW3R017391@repo.freebsd.org> References: <201810132118.w9DLIW3R017391@repo.freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Mateusz,
On Sat, Oct 13, 2018 at 09:18:32PM +0000, Mateusz Guzik wrote:
M> Return is almost always 0. The change replaces 3 branches with 1 in the common
M> case.
This isn't true. For a webserver working with blocking sockets
returning EAGAIN for a very large number of syscalls is normal.
I just dtraced on a random Netflix server and in our case we
get 12% of syscalls with non zero error. But our clients are
special, they request data in small chunks. I believe a regular
web server that serves mostly open ended requests will have a
greater ratio of non-zero returns, up to 50%.
Here is script:
#!/usr/sbin/dtrace -s
fbt::cpu_set_syscall_retval:entry
{
@[args[1]] = count();
}
I would be interested if anybody reports results on a busy
web server running nginx.
So, I doubt that using __predict_true() is an optimisation here.
--
Gleb Smirnoff
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20181015191924.GF1044>