Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 5 Mar 2010 22:14:50 -0600
From:      Kevin Day <toasty@dragondata.com>
To:        freebsd-hackers@freebsd.org
Subject:   ACPI/power implementation causing performance loss with i7/Nehalem turbo boost
Message-ID:  <0ECDEB94-E60E-45C7-98AC-5E948DE4649C@dragondata.com>

next in thread | raw e-mail | index | archive | help

Recently I bumped into something very weird. In some CPU heavy =
workloads, FreeBSD ran faster inside VMware's ESX hypervisor than it did =
running natively on bare metal. Simple pure CPU applications (such as =
"openssl speed") would run 10-30% faster on VMware. This seemed very =
counterintuitive, until I discovered what I believe to be the cause.=20

Intel Nehalem and i5/i7 processors have a feature called "Turbo Boost", =
where the more cores that are inactive (ACPI states C2 or C3) the higher =
the clock rate of the active cores. In some processors increasing the =
clock speed by more than 1ghz. On a hunch, I disabled turbo boost =
(through the BIOS) on our ESX system, and this brought the speeds back =
on par with the bare metal FreeBSD box.

So, it seems that the VMware hypervisor is deactivating cores on the CPU =
when idle, but FreeBSD itself isn't. Is anyone working on giving =
FreeBSD's idle loop/scheduler the ability to go into deeper sleep =
states? It seems this would have more than just a power savings benefit =
now.

Intel documentation on Turbo Boost: =
http://download.intel.com/design/processor/applnots/320354.pdf?iid=3Dtech_=
tb+paper

-- Kevin




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?0ECDEB94-E60E-45C7-98AC-5E948DE4649C>