Messages in this thread | | | Date | Tue, 6 Oct 2009 15:53:11 +0200 | From | Ingo Molnar <> | Subject | Re: [RFC][PATCH 0/9] perf trace: support for general-purpose scripting |
| |
* Peter Zijlstra <peterz@infradead.org> wrote:
> On Tue, 2009-10-06 at 11:09 +0200, Ingo Molnar wrote: > > * Tom Zanussi <tzanussi@gmail.com> wrote: > > > > > Known problems/shortcomings: > > > > > > Probably the biggest problem right now is the sorting hack I added as > > > the last patch. It's just meant as a temporary thing, but is there > > > because tracing scripts in general want to see events in the order > > > they happened i.e. timestamp order. [...] > > > > Btw., have you seen the -M/--multiplex option to perf record? It > > multiplexes all events into a single buffer - making them all ordered. > > (The events are in causal ordering in this case even if there's some TSC > > asynchronity) > > It also wrecks large machines.. [...]
With millions of events per sec, for sure. It doesnt with a few thousand per sec. Right now that's the price of guarantee causality. If you _can_ trust your system-wide TSC then it's not needed - but that's only possible on a very small subset of machines currently.
> [...] I've been thinking about limiting the number of CPUs you can > redirect into a single output stream using the output_fd thing, but > then the inherited stuff makes that very hard. > > And we also need a solution for the inhertited counters, the best > would be the per-cpu inherited things, where we use both cpu and pid, > instead of either. > > In short, -M is nice, but it also has significant down sides, esp. > with machines getting more and more cores.
Yeah.
Ingo
| |