Date: Sat, 25 Apr 2015 10:37:40 -0700 From: Adrian Chadd <adrian@freebsd.org> To: "K. Macy" <kmacy@freebsd.org> Cc: "freebsd-arch@freebsd.org" <freebsd-arch@freebsd.org> Subject: Re: RFC: setting performance_cx_lowest=C2 in -HEAD to avoid lock contention on many-CPU boxes Message-ID: <CAJ-Vmo=6ej6b52yoknP_piL5uJYHm_1AxZh7=6AFGuqMO4sLVg@mail.gmail.com> In-Reply-To: <CAHM0Q_NuvZH0GGs-J9xniyt2PZ0qb_kjOaanVOaCxzD0CVzGYg@mail.gmail.com> References: <CAJ-VmonG%2By5gzoYmer70KAswUorvezcZxRSDsQWj47=jsAZ71w@mail.gmail.com> <CAHM0Q_NuvZH0GGs-J9xniyt2PZ0qb_kjOaanVOaCxzD0CVzGYg@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On 25 April 2015 at 10:18, K. Macy <kmacy@freebsd.org> wrote: > Perhaps use an arbitrary cutoff - say <= 8 cores - where the > cx_lowest=C3. This serialization isn't going to hurt on systems with > more modest core counts. Maybe. I bet it's a function of the idle state entry rate and core count - so maybe at 8 cores it'll hurt but only if it's entering idle at a high rate. Eg, if it's taking a hell of a lot of interrupts but not maxing out the CPU. -adrian
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAJ-Vmo=6ej6b52yoknP_piL5uJYHm_1AxZh7=6AFGuqMO4sLVg>