Re: [Aoetools-discuss] Performance
Brought to you by:
ecashin,
elcapitansam
From: Sam H. <sa...@co...> - 2008-03-20 01:25:01
|
Hi Mike, A few thoughts ... First, upgrade to the latest initiator aoe driver available from http://coraid.com/support/linux/. Many distros haven't kept their aoe drivers in sync with the kernel.org kernels. Even then, we only update the kernel.org driver after a long period of stability in the standalone driver. I wouldn't be surprised to find you're running aoe6-22 which I don't think supported jumbo frames. Have you seen the Jumbo Frame note in the README for vblade? You mention you're not seeing a lot of retransmits so this isn't likely to be the root cause, but you'll likely need to incorporate the documented workaround to achieve optimal throughput using jumbo frames. Do your 8169 based adapters perform auto MDIX? If so, it would be a good test to plug the initiator and target directly into each other to eliminate the switch as a potential bottleneck. Does the output of aoe-stat on the initiator show the vblade with the appropriate interfaces and with a payload size you'd expect based on your MTU settings? (you'll need to fetch the latest standalone driver to get the payload in your aoe-stat output). I'm not exactly sure how O_DIRECT affects throughput. Best to start with the defaults to get your bearing, then refine the setup. HTH! Cheers, Sam |