[Beowulf] Re: [APPL:Scitech] Re: Gigabit switch for XServe cluster (grid) computer (fwd from acaird@umich.edu)

[Eugen Leitl]

----- Forwarded message from Andrew J Caird <acaird at umich.edu> -----

From: Andrew J Caird <acaird at umich.edu>
Date: Wed, 11 May 2005 12:30:43 -0400
To: keith Farnsworth <k.farnsworth at qub.ac.uk>
Cc: Scitech Apple List <scitech at lists.apple.com>
Subject: Re: [APPL:Scitech] Re: Gigabit switch for XServe cluster (grid)
	computer
User-Agent: Mozilla Thunderbird 1.0 (X11/20041206)
Reply-To: acaird at umich.edu


On 05/11/2005 12:16 PM, keith Farnsworth wrote:
>Thanks,
>Yes indeed, I intend a private network with the head node providing the 
>a single link to the outside world on its second ethernet connection 
>(router solution). I am glad the private (cluster only) network can be 
>level 2 unmanaged. University computer support will have nothing to do 
>with me as long as I use Apple kit.
>
>This switch business seems very complicated!


It is complicated.  And expensive.  I agree with everything Jay 
mentioned, and I would add Force10 to the list of vendors.  We have 
high-end switches with 100s of ports with a per-port cost of nearly 
$1000 - not quite as much as a decently equipped XServe, but it's into 
the same order of magnitude.  I'll emphasize what Jay said:

If your application stresses the network, you can have the fastest nodes 
in the world and it can all be worth very little if you have a slow or 
flaky network.

The network is more important than most people realize.  We've tested HP 
gigabit against Force10 gigabit, and the Force10 is more than 50% 
faster.  Just because the words on the spec sheets match up doesn't mean 
you'll necessarily get what you expect.  For big clusters, it's worth 
testing with the gear you're interested in.

Both Cisco and Force10 make nice 48-port stackables with per-port costs 
in the sub-US$200 range that actually perform the way you want and 
expect them to.

Also, keep expansion in mind: clusters look inexpensive because you can 
"always add a node for US$3000", but you need to consider the network in 
there, and that can mean buying a new switch for the next node, and may 
mean introducing an unacceptable physical boundry into your cluster network.

Good luck.
--andrew
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Scitech mailing list      (Scitech at lists.apple.com)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/scitech/eugen%40leitl.org

This email sent to eugen at leitl.org

----- End forwarded message -----
-- 
Eugen* Leitl <a href="http://leitl.org">leitl</a>
______________________________________________________________
ICBM: 48.07100, 11.36820            http://www.leitl.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
Url : http://www.scyld.com/pipermail/beowulf/attachments/20050511/56dce449/attachment.bin