Messages in this thread | | | Date | Sun, 13 Apr 2008 23:51:50 +0200 (CEST) | From | Erik Bosman <> | Subject | Re: [PATCH 2/3] x86: Implement prctl PR_GET_TSC and PR_SET_TSC |
| |
On Sat, 12 Apr 2008, H. Peter Anvin wrote:
> > I also saw no mention about performance impact, which need to be > considered whenever *anything* is proposed to be inserted into a hot > path. It may be (heck, *should be*) that the performance impact isn't > measurable, but that needs to be positively established. > > -hpa >
It took me a while, but I have done some tests om my system with and without my patch applied.
All deamons but klogd/syslogd were shut down andI used the following command:
LMBENCH_SCHED="SINGLE" \ /usr/lib/lmbench/bin/{i686-pc,x86_64}-linux-gnu/lat_ctx -s 0 2
The variation was quite high so I ran it a 1000 times for each configuration.
x86_32, patch applied, 1000x repeated, avg: 1.319, rmse: 0.081 x86_32, patch not applied, 1000x repeated, avg: 1.335, rmse: 0.107
x86_64, patch applied, 1000x repeated, avg: 1.417, rmse: 0.0716 x86_64, patch not applied, 1000x repeated, avg: 1.423, rmse: 0.0745
This is on a core 2 duo E6300.
Erik
| |