Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
/ 2004-01-22 11:01:58 -0500 \ george young: > On Thu, 22 Jan 2004 10:23:43 -0500 > Todd Denniston <Todd.Denniston at ssa.crane.navy.mil> threw this fish to the penguins: > > > Hi, > > [drbd-0.6.10, Fedora Core release 1, 2 node] > > As I was beginning to create my drbd.conf, using the > > /usr/share/doc/drbd-0.6.10/drbd.conf example, I read the following comment: > > > > # btw, don't do this. > > # did you notice that in this example we have two drbd devices > > # on the same spindle (hda)? performance will be bad. if you > > # use several drbd devices, put them on different spindles; > > # different channels/controllers won't be a bad idea for IDE. > > > > I have a raid box [Promise RM8K, RAID 5, ~750GB, SCSI], which presents itself > > as one disk, AKA spindle, and I wanted to know if the above note was mainly now, if it has more than one physical spindle, it _has_ more than one spindle. So the argument does not hold exactly. > > for [IDE|single physical spindle] or if it also badly affects RAID boxes > > presenting as one disk? > > > > Is this because DRBD/EXT3 makes some assumptions about stacking up writes that > > are not true when presented with multiple partitions, or just that the seek > > times get amplified badly with drbd in protocol=C? it is mostly because of seektime killing your resynchronization performance. you can avoid that nowadays by putting the devices in different "sync-group"s. > > I wanted to split the space up into partitions such that business functions > > [cvs & project storage] would be less troubled by disk hogs home directories. > > > > can or is it wise, to have drbd mirror the whole device sda and have > > partitions in it? you cannot have partitions in drbd. what you can do is have drbd as a LVM2-PV. but you should know why you are doing it before you try this. > Some RAID controllers allow you to configure multiple "virtual drives", > i.e. if the array has enough physical drives, split them into two or > three groups, each group a raid 5 or whatever. This would avoid most of > the "same spindle" performance problem. You still have a possible > bottleneck at the raid ctlr itself it it's not fast enough at calculating > ecc data for writes. > > Also keep in mind that you are mirroring a block device, not a file system. > DRBD doesn't know or care that you may only be using 20% of a given fs, > it will still sync the whole device's contents across the net when it > chooses. Work out how long it would take to copy your entire 750GB > across your net... Yes. Don't mirror more than neccessary. I'd suggest to have a LVM logical volume as the lower level device for drbd. If you later figure you need more space, you can easily but *carefully* do lvextend, change the disk-size in drbd.conf, e2fsck -f, resize2fs, and be happy... Lars Ellenberg