Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Thu, 29 Mar 2007, Duane Cox wrote: > Follow up: > > I was hoping Lars or Phillip would provide some pointers, but I see that > other various people have reported similiar problems an go unanswered.... > > I have been running this command to benchmark WRITE performance to the SCSI > block device. > The funny thing is that each additional run, the WRITE performace > decreases... INTERESTING.... > > > ./dm -a 0 -b 1M -s 3G -y -m -p -o /dev/sdb > 71.87 MB/sec (3221225472 B / 00:42.745236) I'm sorry, I should have noticed this earlier. The test is too short. Can you be sure there is no caching in effect? A longer test helps to eliminate that. Try running it longer, like 10x as long. Also, just to test it, try making a single-disk software raid1 out of /dev/sdb and telling drbd to use /dev/mdwhatever as it's backing store. I've seen wacky performance differences depending on the "most close" block device underlying a given store. Easily demonstrated with raw device access with aoe (aoe will use 512b frames if you just dd to the device, whereas if you use md on top of the aoe device you'll get multiples of 512b frames). -- Jon Nelson <jnelson-drbd at jamponi.net>