Jumbo Frames

Posted: 2010-10-01 18:35:44 by Alasdair Keyes

Direct Link |

We'd been having some issues with retransmit errors on our iSCSI network, so we enabled jumbo frames to allow frames up to 9000 bytes instead of the standard 1500. The change was very easy to implement...

ifconfig ethx mtu 9000

And then updated the /etc/sysconfig/network-scripts/ifcfg-ethx file to include the line

MTU=9000

(We use CentOS)

Using the Dell SANHQ software showed a marked increase in the speed of large iSCSI transactions, However we discovered that some of our CentOS based Xen hosts were dropping incoming packets on the iSCSI interface

$ ifconfig ethx | grep dropped
          RX packets:1883666 errors:0 dropped:12636 overruns:0 frame:0
          TX packets:1696267 errors:0 dropped:0 overruns:0 carrier:0

This was a concern as we'd not had any issues with this before, although some dropped packets can be forgiven, in this case we seemed to have problems with our Hosts kernel panicing and rebooting. We couldn't work out what was causing these dropped packets and tried a number of things including increasing the incoming TCP buffer, but to no avail.

After a short time I noticed that the hosts running slightly earlier kernels were getting the packet loss, whereas the later ones were fine, specifically Redhat/CentOS kernels 2.6.18-164.x.x.el5xen were dropping packets, where as the 2.6.18-194.x.x.el5xen kernels were fine.

There's no special fix in this instance, if you're getting s similar issue and running the 2.6.18-164 kernel make sure you do an update. It's also just a reminder to keep your servers and especially kernels, up to date, although software is often only updated for security patches, there are many small fixes that are rolled out that fix little issues like this.

If you found this useful, please feel free to donate via bitcoin to 1NT2ErDzLDBPB8CDLk6j1qUdT6FmxkMmNz