On 01/11/2011 02:08 AM, Kinkie wrote: > On Tue, Jan 11, 2011 at 8:03 AM, Amos Jeffries <[email protected]> wrote: >> The profiler aufs assertions are caused by the profiler not being thread >> safe but attempting to account operations inside each of the AIO threads. >> Lack of thread safety is due to the stack the profiler maintains of what >> states are currently being counted. >> >> I don't believe we should be maintaining such a stack. Instead I think we >> should leverage the existing system stack by having the PROF_start(X) macro >> create a counter object in the local scope being counted. When the local >> scope exists for any reason the system stack object will be destructed and >> the destructor can accumulate the counters back into the global data. > > +1. I like this.
I have a crude patch that implements such scope-based profiling. We used it to find performance regressions in Squid3 (still an ongoing project)... I would be happy to find that patch and post if somebody wants to polish it. Thank you, Alex.
