[DRBD-user] write performance issues

Mrten mrten+drbd at ii.nl
Mon Jul 11 17:05:10 CEST 2011

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


On 11-7-2011 16:23, Mark Dokter wrote:

> I had the syncer rate at 30M before and set it to that high value out
> of desperation. There's nothing much going on on the syncer network,
> so this shouldn't be a problem.

Depending on the protocol used (didn't see it in your config, I'm
guessing C is default?) you'd want as low as a latency as possible.

If protocol C, writes aren't "done" on primary if not confirmed by
secondary.

Roundtrip times then add up.

> The two servers both have a dedicated nics which are connected to a
> gbit switch. Some time ago, those two servers were right next to each
> other (physically) and I tried a direct connection of those two nics,
> which did not really make a difference, so I plugged them to the
> switch again. Now those two servers are physically about 1km appart
> in two different server facilities and connected via VLAN. That
> shouldn't really matter, as bandwith and latency between those two
> servers is quite acceptable (ping measures ~0.2 ms).

My dedicated (bonded) links measure 0.007 ms. You could try another
protocol, see if that helps.

But my guess is that 'lots of small writes' makes one of the raid-disks
bottleneck for activity log updates (which causes head-seeks, which
causes slowness). Can you tune the HW-RAID to cache more?

Try another physical device for the metadata, helped tremendously in my
case (70 > 250 MB/s with dd).

Is your filesystem (and is your LVM) aligned with the chunksize on the
HW-RAID? Might cause a read/write/write otherwise, could be extra
detrimental on raid-6.

Lots of weasel words there, sorry!

Mrten.



More information about the drbd-user mailing list