On Thu, Nov 17, 2011 at 8:22 AM, Jeff Trawick <[email protected]> wrote: > On Thu, Nov 17, 2011 at 8:00 AM, Jim Jagielski <[email protected]> wrote: >> >> On Nov 17, 2011, at 7:32 AM, Jeff Trawick wrote: >>> >>> FWIW, using the optimized rough time from the Event listener thread >>> within mod_reqtimeout is consistently faster in my testing. >> >> Same here... > > Thanks for trying it out; my individual runs of a particular testcase > were not consistent enough to get abundant warm fuzzies but it did > look good considering the ranges of runs with/without the change. > > In the next couple of days I'll try to actually *read* the listener > thread code (to see if the optimization needs to be disabled during > certain intervals, like shutdown) and clean up/commit the API to > trunk. As long as it remains an API that works even if the MPM > doesn't/can't-practically provide an optimized version, it can go into > 2.4.x at any point.
There's also the atomicity of load/store in the event implementation to consider... 64-bit apr_time_t, no portable 64-bit atomics in APR... it could be bad news if a caller sees the new high-order 32-bits and the old low-order 32-bits...
