[DRBD-user] Re: DRBD doesn't seem to scale to fast underlying hardware with typical database load

[Lars Ellenberg]

On Sat, Aug 09, 2008 at 11:44:25AM -0700, Casey Allen Shobe wrote:
> On Aug 9, 2008, at 11:21 AM, Casey Allen Shobe wrote:
> 
>     * Maybe JFS is a bad choice to use on top of DRBD (kind of doubt it) and we
>     should use another filesystem who's disk accesses match DRBD's expectations
>     better.
> 
> 
> After sending this E-mail, I decided to mkfs.ext3 the partition (with DRBD
> slave connected).  It got to the "writing inode tables" part, shot up to 1200
> pretty fast and then stuck.  It's now been CRAWLING along at a snail's pace for
> a good 10 minutes now, and is not yet up to 2000 (out of 5465, so I expect it
> will be the better part of an hour or more before it completes, which seems
> absolutely absurd for an mkfs).  Additionally, I tried disconnecting the slave
> midway through and speed did not improve.
> 
> So the problem seems lower-level than the filesystem (just wanted to rule that
> out).
> 
> It also seems, since performance is terrible even when the slave is
> disconnected, that the network is not a factor.
> 
> The mkfs alone is pushing the server load up to 3.0, however the CPU
> utilization is sticking at 0%, so CPU does not seem to be a factor either.

familiarize yourself with the concept of the DRBD activity log.
the config parameter would be "al-extents".

verify that your disk subsystem has a non-volatile write cache enabled.
then add to DRBDs configuration no-disk-flushes and no-md-flushes.

see if that helps.

-- 
: Lars Ellenberg                           http://www.linbit.com :
: DRBD/HA support and consulting             sales at linbit.com :
: LINBIT Information Technologies GmbH      Tel +43-1-8178292-0  :
: Vivenotgasse 48, A-1120 Vienna/Europe     Fax +43-1-8178292-82 :
__
please don't Cc me, but send to list -- I'm subscribed