[DRBD-user] Optimization Issues

Lars Ellenberg lars.ellenberg at linbit.com
Fri May 7 15:15:01 CEST 2004

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


On Fri, May 07, 2004 at 12:04:27PM +0800, Federico Sevilla III wrote:
> Hi everyone,
> 
> I have a two-node HA cluster using DRBD for data replication.
> Filesystems are all ext3 in ordered-write mode, DRBD v0.6.12, Linux
> 2.4.26 and Debian GNU/Linux 3.0r2 with DRBD taken direct from
> <http://fsrc.csee.wvu.edu/debian/apt-repository>.
> 
> The two machines are connected to each other via GbE with a crossover
> cable. They have two physical hardware SCSI RAID 10 arrays each, and two
> DRBD devices set up, one for each array. One DRBD device is low-traffic,
> used for shared configs. The other DRBD device is used heavily for the
> PostgreSQL database.
> 
> During synchronization and when there is heavy database I/O, I get a lot
> of the following messages in my logs on the primary node (interestingly,
> the secondary node doesn't complain at all):
> 
>     kernel: drbd1: transferlog too small!!
>     kernel: drbd1: tl messed up!
>     kernel: drbd1: Epoch set size wrong!!found=192 reported=191
> 
> I searched the archives and found that this basically means I need to
> tune my drbd.conf file, but isn't something critical. Is this correct?
> Would anyone know of a general tuning and optimization guide for DRBD?
> Or perhaps would anyone be able to spare me some time to comment on my
> DRBD configuration file?
> 
> I'm also curious: what's the most reliable way of finding the value to
> put in disk-size? Or can this be omitted for configurations where the
> partitions on both sides are of exactly the same size?
> 
> My configuration is as follows:
> 
>     resource drbd0 {
>         protocol = C
>         fsckcmd = fsck -p -y
>         inittimeout = 60
>         disk {
>              do-panic
>              disk-size = 61522304k
>         }
>         net {
>              sndbuf-size = 1M

note that increasing the sndbuf-size does not neccessarily improve
performance. experiment with this value, maybe even decreasing it helps
performance!

>              sync-nice = -20
>              sync-min = 4M
>              sync-max = 600M
>              tl-size = 5000

well, if DRBD complains about tl too small, up this value.
say, 1500, 8000, 12000000 :) until that message no longer triggers.

for further "optimization" you might try to use "jumbo frames" on your
nic, mtu 5000 or some such.



	Lars Ellenberg



More information about the drbd-user mailing list