Huge pages, slow drives, and long delays [LWN.net]

Huge pages, slow drives, and long delays

Posted Nov 18, 2011 13:39 UTC (Fri) by foom (subscriber, #14868)
Parent article: Huge pages, slow drives, and long delays

Awesome timing on this article, this is almost certainly exactly what's slowing down a big server workload I run when transparent hugepages are enabled.

We tried experimentally enabling the THP feature in the latest kernel but it made things *much* slower, which seemed really mysterious. Now, knowing that the kernel is unnecessarily blocking processes on writing some pages to disk explains the whole issue. (Note: no USB sticks here; normal hard drives are already slow enough for me...).

So yeah, that argument about how it's desirable for long-running server workloads? Not so much.

(Log in to post comments)

Huge pages, slow drives, and long delays

Posted Nov 19, 2011 20:56 UTC (Sat) by Ben_P (subscriber, #74247) [Link]

Are long running server workloads considered those who hold large chunks of memory for a long time? Almost any Java application would fall under this due to it's own internal memory management. It does seem very plausible there are some server loads that actually malloc/free per-request or something similar which could make this problem seem worse.