On Wed, 21 Sep 2011, Clark Williams wrote:

While kernel.org is sorting out the security stuff, the rt-tests code
may be pulled from:

	git://github.com/clrkwllms/rt-tests.git

Note that we're now at version 0.82. Presently I only have the git
archive available (no tarballs). 

While investigating latency spikes in the 3.0.x-rt kernels, Thomas
spotted a case where an Intel quad-core Xeon was going into deep
sleep states and were all fighting to come out of sleep at the same
time (and consequently causing a big latency spike in cyclictest). 

While trying to figure out how to prevent deep cstates I remembered a
conversation I had with Arjan at the last Plumbers conference in
Boston. He mentioned the /dev/cpu_dma_latency interface to the power
managment code and that if you opened it and wrote a zero to it, you
effectively put the system into "idle=poll" mode until you closed the
file descriptor (see: Documentation/power/pm_qos_interface.txt). 

I've added a set_latency_target() function to cyclictest that by
default opens /dev/cpu_dma_latency and writes a zero to it, then holds
the file descriptor open for the duration of the cyclictest run. This
made a *huge* difference on some Intel Xeon's. Without this option, when
I was running cyclictest with the -b option, I saw latencies over
300us. When I added it, while tracing I never saw a latency over 30us.
Turning of -b, I never saw it go over 10us. I am doing further testing
now with other x86_64 systems. 

Of course this is very architecture specific, so YMMV, but I think it's
a valid mechanism to be used when measuring latency and I believe a
technique that many latency-sensitive applications might use to good
effect. 

This is not a big surprise as the kernel got more agressive going into
deep C-states since 2.6.33-rt especially with "intel_idle" on modern
(Nehalem+) cpus.

Thanks,

	tglx