Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Wed, Jul 16, 2008 at 07:12:56PM +0200, Lars Marowsky-Bree wrote: > On 2008-07-16T12:05:53, nathan at robotics.net wrote: > > >> You're being hit by the internal metadata. Putting the metadata on > >> another spindle will make it quite fast. > > What is the metadata overhead in DRBD? I am seeing about 100 MB/s overhead > > with DRBD: > > It's writing to the activity log when you dirty a new block; that costs > you latency, because the head has to move from the data area to the > metadata block. > > And the latency indirectly affects the bandwidth you get. > > Putting the metadata on a different, independent avoids this. SSDs are > really cool for this, as the metadata is also very small (just 128MB per > device). well. you should get yourself a decent BBU non-volatile Cache on your controller, and be done with it. btw, for Nathan's benchmark (using dd if=/dev/drbd of=/dev/null), it is _reading_ only, which does not (yet) change any drbd meta data. my guess is that Nathan sees some caching effects, either from page cache, or from the controller cache. for streaming read performance, make sure you benchmark cache cold always. it may also be a queue depth or read-ahead or whatever effect. whatever it is, read performance on DRBD should always be practically indistinguishable from read perfromance on the lower level device. if it is not, something is wrong (not in DRBD, but in your setup). for writes it is much more more involved. -- : 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