On Fri, 9 May 2014, Christoph Lameter wrote: > On Fri, 9 May 2014, Thomas Gleixner wrote: > > I think we agreed long ago, that for the whole HPC FULL_NOHZ stuff you > > have to sacrify at least one CPU for housekeeping purposes of all > > kinds, timekeeping, statistics and whatever. > > Ok how do I figure out that cpu? I'd rather have a specific cpu that > never changes. I followed the full nohz development only losely, but back then when all started here at my place with frederic, we had a way to define the housekeeper cpu. I think we lazily had it hardwired to 0 :) That probably changed, but I'm sure there is still a way to define a housekeeper. And we should simply force the timekeeping on that housekeeper. That comes with the price, that the housekeeper is not allowed to go deep idle, but I bet that in HPC scenarios this does not matter at all simply because the whole machine is under full load. Frederic? > > So if you have a housekeeper, then it makes absolutely no sense at all > > to move it around in circles. > > > > Can you please enlighten me why we need this at all? > > The vmstat kworker thread checks every 2 seconds if there are vmstat > updates that need to be folded into the global statistics. This is not > necessary if the application is running and no OS services are being used. > Thus we could switch off vmstat updates and avoid taking the processor > away from the application. > > This has also been noted by multiple other people at was brought up at the > mm summit by others who noted the same issues. I understand why you want to get this done by a housekeeper, I just did not understand why we need this whole move it around business is required. Thanks, tglx -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>