vMotion & Jumbo frame

Anyone enabled jumbo frame for vMotion? There was an article I read while back which recommended not enabling jumbo for vMotion due to mixed performance. Just finished watching a presentation by VCDX recommending jumbo for vMotion (he did state for 10 Gig). Curious as to how TE members have their environment setup.

Find more posts tagged with

Comments

MickQ

We've done it and are having problems. Not sure if it's the network card drivers or what exactly.

blargoe

We're not doing it currently, I never have.

jibbajabba

MickQ wrote: »

We've done it and are having problems. Not sure if it's the network card drivers or what exactly.

Make sure the switches have them enabled too. Cisco switches for example don't have the jumbo frames enabled by default.

MickQ

Have done that. There seems to be other quirks in the HP NIC drivers interfering with it all - failover problems, etc.

ptilsen

We had problems on multipathed GbE iSCSI network, but never tried in on LAN or any 10G networking. It caused some issues on the iSCSI network, though they might not have been VMware specific. I would imagine it is a significant benefit on 10G.

/usr

We have jumbo frames enabled for vMotion. Have not noticed any performance issues related to vMotion which would lead me to believe that jumbo frames were the culprit. Actually, prior to turning on jumbo frames on the vSwitch, I was getting failed vMotions on two newly installed ESXi hosts.

What issues have you guys seen?

jasonboche

Speaking from an IP storage fabric perspective, enabling jumbo frames is commonly parroted as a best practice throughout the industry but I don't think it's a one size fits all design decision. There are risks associated with it which can lead to downtime. There is a hidden cost in terms of complexity. Enabling Jumbo Frames does not yield order of magnitiude performance gains. Throughput gains are closer to marginal. Narrowing the discussion to a vMotion context, risks involved are somewhat isolated but my own testing with 1Gb Ethernet did not reveal much performance gain. Others with access to 10Gb lab environments showed consistent results in minimal gains (0-15%). Weigh those gains against risk and complexity. Does it still make sense? Maybe in a bet-the-business situation, but I wouldn't recommend it globally. Those that make wild accusations like "jumbo frames increases performance 2x-9x!" haven't tested themselves & I wouldn't consider a credible source.