Messages in this thread | | | Date | Mon, 19 Jun 2017 16:52:05 +0100 | From | Mark Rutland <> | Subject | Re: [PATCH v3 1/n] perf/core: addressing 4x slowdown during per-process profiling of STREAM benchmark on Intel Xeon Phi |
| |
On Mon, Jun 19, 2017 at 08:39:18AM -0700, Andi Kleen wrote: > On Mon, Jun 19, 2017 at 04:24:01PM +0100, Mark Rutland wrote: > > On Mon, Jun 19, 2017 at 08:21:51AM -0700, Andi Kleen wrote: > > > > I was trying to get a feel for how that compares to what we can do > > > > today. For other reasons (e.g. fd exhaustion), opening NR_CPUS * n > > > > > > You just have to increase the fd limit. The 1024 fd default is just > > > archaic for larger systems and doesn't really make any sense because > > > it only controls very small amounts of kernel memory. > > > > > > > events might not be a great idea on systems with a huge number of CPUs. > > > > We might want a heuristic in the perf tool regardless. > > > > > > But there's no alternative: we have to measure all CPUs with all events. > > > > You can measure the process on all CPUs by using 1 event without a CPU > > filter, rather than NR_CPUS events. > > That wouldn't measure all threads, at least not with current perf core.
Ah; I missed the constraint imposed by perf_mmap().
For some reason I thought that was enforced by userspace only.
Sorry for the noise.
Mark.
| |