[Beowulf] Intel CSA performance?

Robert Myers rmyers1400 at comcast.net
Sun Mar 21 01:40:30 EST 2004


Greg Lindahl wrote:

>On Fri, Mar 19, 2004 at 10:58:53PM -0500, Robert Myers wrote:
>
>  
>
>>I had thought it might be interesting to fool around with trying to use 
>>CSA for hyperscsi, but I think you're saying if you're going to use a 
>>switched network, don't bother, if you're trying to win on latency.
>>    
>>
>
>I've never heard of "hyperscsi", and I am not saying what you think
>I'm saying. What I am saying is that if you're going to use 1 gigabit
>Ethernet, which has high latency in the switches, AND go through the
>kernel, don't bother. I was pretty clear, so I don't see how you
>missed it. There are certainly many examples of switched networks that
>are low latency, such as Myrinet, IB, Quadrics, SCI, and so forth.
>
I should have been explicit. "If you are going through a switched 
_ethernet_ connection." If you do the groups.google.com search

low-latency infiniband group:comp.arch author:Robert author:Myers

you will find that you really don't need to educate me about the 
existence of low-latency interconnects.

As to hyperscsi, I gather that it is incumbent only on others to check 
google. Hyperscsi is a way to pass raw data over ethernet without going 
through the TCP/IP stack:

http://www.linuxdevices.com/files/misc/hyperscsi.pdf

so it doesn't consume nearly the CPU resources that TCP/IP does without 
hardware offload, and I don't think CSA allows you to use separate 
hardware TCP/IP offload. It looks potentially interesting as a low-cost 
clustering interconnect, especially if, as I expect, Intel continues to 
push ethernet.

RM




_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list