On Tue, Jun 22, 2010 at 8:59 PM, Guido Trotter <[email protected]> wrote: > > Was thinking the same. Will do. >
Ok, I have results. I just ran 40 jobs, with 5 opcodes each, each logging 5 times (as it was slow enough) and sleeping for 0 time. As expected in this case we lose, as the data gets serialized and written to a file while holding the jqueue lock in an exclusive manner. This will improve as we tackle this lock more, though: if you prefer, and the patch concept is ok, I can avoid submitting these patches until I have better numbers (which I'm sure to achieve) but keep the LGTMs on the current series and build on top of it. Current master (+ logging delay patch): C:op 200 C:job 40 T:submit 0.2283 T:exec 13.7781 T:total 14.0064 With just writing to file enabled: C:op 200 C:job 40 T:submit 0.2282 T:exec 30.6113 T:total 30.8395 With everything I posted (condition removal and watching via inotify): C:op 200 C:job 40 T:submit 0.2339 T:exec 29.6945 T:total 29.9283
