On Tue, 6 Aug 2013, Alex Thorlton wrote: > I've gotten my hands on some of the benchmarks/code that were used to > originally uncover the performance issues we're seeing. I'm currently > trying to separate out the performance issues that are being caused by > the kernel code from issues involving hardware - the cost of remote > memory accesses is a bit higher on our systems with node controllers vs. > glueless QPI/Hypertransport-based systems. >
We've seen some issues where accessing remote hugepages causes performance degradations over accessing local pages that affects some workloads, but that seems like a mempolicy issue rather than madvise. You probably want local hugepages but not fallback to remote hugepages before allocating local pages? It would be interesting to see /proc/pid/numa_maps and /proc/pid/smaps for workloads that don't like thp. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
