Messages in this thread | | | Date | Fri, 22 May 2009 20:07:45 +1000 | From | Paul Mackerras <> | Subject | Re: [PATCH 2/2] perf_counter: optimize context switch between identical inherited contexts |
| |
Peter Zijlstra writes:
> Ah, should I read this as 'n' lmbench instances on a single cpu? So that > we get multiple inheritance sets mixed and have to switch between them?
No, it's one lat_ctx instance, which creates N processes connected in a ring with pipes, and passes a token around some number of times and measures the overall time to do that. It subtracts off the time it takes to do all the pipe reads and writes in a single process to get an estimate of how much the context switches are costing.
Paul.
| |