Re: [Aoetools-discuss] Vblades - how many possible
Brought to you by:
ecashin,
elcapitansam
From: kelsey h. <kh...@dr...> - 2008-02-05 15:42:40
|
Sam Hopkins wrote: >>> Upgrade to the latest driver available on our web site: >> If this driver doesn't somehow resolve the problem of an incoming packet >> waking up all 1500 of his proposed vblade processes I can only imagine >> that performance will seriously suck for all of those context switches >> happening for every packet. That is the main reason I am looking at >> other implementations of the aoe target. > > That's not an initiator driver problem. That's a userspace raw > packets problem. ... caused by the vblade process being unable to service more than one target at a time. Yes, pedantically it's a userspace raw socket io problem, but keep in mind that this particular userspace raw socket io problem is due to a target implementation that is, while simple, designed in such a way that it overlooks this problem. So yes, it is a target implementation problem. I think the point that Tracy was trying to make is that a new version of the driver is unlikely to solve his performance problems given that so many discrete vblade processes will be running. I've moved to qaoed on all my AoE targets. Although it's got some minor packet addressing quirks (that don't affect proper target or initiator operation), it seems to be stable. I haven't had any major issues with it. On my systems which were previously running 35 vblade processes, this both increased throughput dramatically while decreasing load on the server; the load averages dropped from 3-4 to less than 1, and the amount of time spent in softirq (context switching) was vastly decreased to near nil. -kelsey |