Re: Importing iSCSI target from NetBSD

On Tue, May 30, 2006 at 09:46:39AM +0100, Robert Watson wrote:
> 
> On Mon, 29 May 2006, Brooks Davis wrote:
> 
> >On Mon, May 29, 2006 at 05:43:28PM -0700, Paul Allen wrote:
> >>>From Scott Long <scottl_at_samsco.org>, Mon, May 29, 2006 at 05:51:28PM 
> >>>-0600:
> >>>>P.S. Just to make it clear - just consider running iSCSI over 100MBps
> >>>>link or even a slower WAN links, which I think covers very large market
> >>>>for this technology now. Performance constrain imposed by running in
> >>>>userland is unlikely to be an issue at all.
> >>>
> >>>Every company and group that I've talked to about iSCSI is worried about 
> >>>performance.  In any case, please follow the lead of Mr. Senault and 
> >>>look at making this a port.
> >>
> >>And in particular the anticipation of low(er) cost 10Gb Ethernet is a 
> >>driving factor behind iSCSI.
> >>
> >>AFAIK, the low-latency performer in this field (for NICs) is from 
> >>Myricom. Andrew Gallatin (one of the FreeBSD alpha committers)  was 
> >>responsible for porting the myrinet drivers, so perhaps he can comment as 
> >>to whether FreeBSD will be getting a driver for their 10GbE cards.  
> >>Ethernet at these speeds is real stress-test for many OSs; it should be 
> >>interesting to see how FreeBSD holds-up.
> >
> >There's a driver in current.  We don't perform nearly as well as we should 
> >at the moment.
> 
> FYI, I recently received donated hardware from Yahoo! and Drew has kindly 
> offered to send me a couple of 10gbps cards to work with, so I hope to have 
> a chance to start doing some measurement and optimization work.  One of the 
> problems we've been having is that it's hard to optimize the CPU use of the 
> network stack when the CPU significantly outstrips available bus and 
> network bandwidth.  It seems like hardware swings back and forth quite a 
> bit -- for a few years gigabit was way-the-heck-faster-than-CPU, now it's 
> the other way around again.  The best stack optimization work happens when 
> you have to figure out how to get the network stack to perform well in 
> near-infinite bandwidth scenarios with a CPU-bound stack, which is where we 
> are with 10gbps currently.  One of the things that makes all this rather 
> tricky is that it's quite hard to build test rigs, test setups, and get the 
> hardware details right.  Hopefully, with Yahoo's and Drew's help, my test 
> setup will be good for looking at this for a couple of years.

That's great news.  This problem of relative performance leapfrogging is
a pain.

-- Brooks

-- 
Any statement of the form "X is the one, true Y" is FALSE.
PGP fingerprint 655D 519C 26A7 82E7 2529  9BF0 5D8E 8BE9 F238 1AD4