Fitzcarraldo wrote:DontPanic, how did you get on with the Mathieu Desnoyers patches? Any noticeable difference?
I've had them installed on two machines for about two weeks now. I haven't really noticed any difference with them activated.
The "vmscan: raise the bar to PAGEOUT_IO_SYNC stalls" patch gave me a much more noticeable improvement.
I've been trying to benchmark the Mathieu Desnoyers/Nokia/Efficios patches with Con Kolivas' latency benchmark, interbench (I've added an ebuild to the overlay for this benchmark), and testing the same kernel with and without the patches enabled. This benchmark produces a lot of output that does not summarize very well. Also, some of the results are difficult to reproduce exactly due to the very nature of trying to measure latency when injecting randomness into the system.
But out of a battery of 32 latency tests, I'd say the patches had poorer latency performace (as measured by this benchmarking tool) about two-thirds of the time for those tests that showed a clear difference.
So, for my part, I've more or less lost interest in further work with this patch set. I don't notice any qualitative or quantitative difference, and the approach seemed to be meeting some resistance based on the feedback in the lkml thread discussing these patches.
Let me know if you have run across a better benchmark for measuring latency, and I'll try it.