[DRBD-user] DRBD serious locking due to TOE - RESULT - UPDATE
Jure Pečar
pegasus at nerv.eu.org
Sat Dec 22 00:52:41 CET 2007
Coming late in this thread, I too (just recently) identified problems with
drbd on HP DL385 machines with Broadcom NICs.
The problem is easily repeatable even on minimal install: just enable xinetd
chargen on one machine and suck data from it from the other (something like
nc <ip> 19 > /dev/null) and watch iptraf -d on that interface on sending
machine. IP checksum error counter goes up at about 10% rate of the packet
count.
The thing one has to turn off with ethtool is not ip checksum, but tcp
segmentation (ethtool -K eth1 tso off). This is what is really making
problems. I confirmed it on 1.4.43, 1.4.52d and 1.5.11-rh version of the
driver. In january, when I come back to work, I'll repeat tests with 1.6.7
to see if it makes any difference.
IMHO it's a shame that neither Broadcom nor HP/Dell did not find such
problem in their QA tests and released such broken NICs to the market.
--
Jure Pečar
http://jure.pecar.org/
More information about the drbd-user
mailing list