On Mon, Sep 24, 2018 at 10:23:37PM +0300, Alexey Budankov wrote: SNIP
> > Workload: matrix multiplication in 128 threads > > > > P (period, ms) : 3 > > runtime overhead (%) : 1.8x ~ 12.58 / 6.81 > > data loss (%) : 9 > > LOST events : 147 > > SAMPLE events : 673299 > > perf.data size (GiB) : 0.8 > > Please see more comparable data by P (period, ms), > runtime overhead and data loss metrics at the same time. > > It start from serial implementation as the baseline and > then demonstrates possible improvement applying configurable > --aio(=N) and --threads(=T) implementations. > > Smaller P values, with data loss and runtime overhead values > equal or in small vicinity of the ones from serial implementation, > might mean possible gain. sry for delay.. ok, so it's not so bad afterall ;-) thanks a lot for running the test I need to rewrite some parts of it for the next post, but I'd hate to lose your aio implementation and the possibility to easily compare it against threaded implementation I think we are able to keep it along under --aio option together with current (sync) write implementation and future threads implementation.. could you make it available only under --aio option (or such) and repost? thanks, jirka
