[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

data=journal with large external journals



I have a heavily loaded Apache 2 server which is experiencing what appear to be "write storms". The server is a dual Xeon box with hyperthreading, and appears to linux as a four-cpu box. It has 4GB of physical RAM and is never hitting swap.

It's serving a large number of static image files off of four 135GB SCSI drives with external journals on a fifth volume. The journal volume is an identical disk to the other four, split into four equal journal volumes. The data volumes are mounted noatime,data=journal. This configuration was chosen because users are uploading images to the server and I thought that putting writes on an external volume would improve read efficiency.

I'm running kernel 2.4.21 with Andrea Arcangeli's rc8aa1 patch set (mostly for the elevator-lowlatency patches which fix a problem with huge amounts of cpu being used by kswapd). This patch set also includes the akpm jbd_refile_buffer patch to transaction.c.

The symptoms are that the system chugs along for hours until the load suddenly shoots through the ceiling (500+) as the buffer utilization drops to nearly zero, rendering the box mostly unresponsive for several minutes. What I suspect is happening is that my huge journals are filling up without being flushed to disk, and then ext3 is flushing the entire journal.

My current bdflush settings are

40 500 0 0 60 300 60 0 0

as recommended by others on this list and elsewhere. It doesn't seem to help.

I'm a little stumped at the moment. Any hints would be greatly appreciated.

--
Kevin Murphy
kevin eorbit net








[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]