[Beowulf] Tyan's Beowulf in a box

[Vincent Diepeveen]

Yes i had received around those release dates already quite some questions 
whether my software works on this box.

Tyan is hammering upon their 'direct connect' architecture.

At home here i've got a stand alone S2881 mainboard, it just has 2 pci-x 
133Mhz slots.

It's a pretty good mainboard except that the built in graphics card is 
really too slow nowadays for such a mainboard,
even though i'm not doing much graphical at it other than now and then a 
'dir' in a command prompt.

Having pci-x then is a bad idea.

It would be pretty bad of course if this 'personal supercomputer' is just a 
case with 4 stand alone mainboards without
any form of connection between the boards.

Note that if you want to buy a real 8 socket machine, tyan has a different 
alternative that's more capable of delivering 8 sockets.
It's pretty cheap too.

It's 2 quad boards that are glued together with interconnects (seems to me 
like htx slots by the way), and it works pretty great
and is already operational here and there.

Behaves as 1 big 8 socket machine. Of course latency from each chip to the 
global shared memory is pretty bad as compared to such a S2881 board and the 
cpu's for 8xx are more expensive than the 2xx dual core chips.

That 8 socket machine gets powered then with in total 1300 watt PSU.

Back on the typhoon 'personal supercomputer'; using gigabit would be a 
scandal for such a capable box.
Using 350 watt PSU's is quite risky IMHO if dual cores get put in and some 
i/o happens.

The TDP of those dual cores is already 95 watt and up by now.

That's 180 watt, excluding the watts the PSU will lose for sure, just for 
the chips. Add to that some 4 SATA drives.
Each drive if it's a maxtor will be > 22 watt. Then we already are way above 
300 watt at least for 1 psu.

Then usually the financial companies buying toys like this put in a lot of 
ram, as that what's what they buy a cluster for,
just the RAM they want to use for usually 1 big memory mapped file.

That'll eat quite a lot of power too. 350 watt a board is not exactly what 
i'd advice.

4 x 350 = 1400 watt, without spreading risks over each power supply.

What if 1 of those psu's fails?

Vincent

----- Original Message ----- 
From: "Andrew Piskorski" <atp at piskorski.com>
To: <beowulf at beowulf.org>
Sent: Friday, June 09, 2006 10:19 PM
Subject: Re: [Beowulf] Tyan's Beowulf in a box


> On Thu, Jun 08, 2006 at 11:23:00PM -0500, Geoff Jacobs wrote:
>> Found this on the Tech Report:
>> http://www.tyan.com/PRODUCTS/html/typhoon_b2881_spec.html
>>
>> Looks like a fairly slick setup. Basically, you get a four node beowulf
>> in a box the size of a DS20, suitable for deskside use. The only problem
>> is that, from what I can see, it only supports GbE.
>
> I thought this came up before on the list when Tyan first announced
> it, including pictures of early models at the show, some discussion of
> the similar (but perhaps superior) boxes from Rocketcalc and Penguin
> which have been around much longer, etc.  Perhaps that was elsewhere
> though, as for some reason I can't seem to find it in the list
> archives.
>
> Looks like Tyan announced this Typhoon box back in March:
>
>  http://www.tyan.com/html/pr06_typhoon_psc.html
>  http://www.tyan.com/products/html/typhoon_b2881.html
>
> Tyan's Typoon is using their S2881 motherboard.  Interestingly,
> Rocketcalc says they use either the same Tyan S2881, or the S2891:
>
>  http://www.rocketcalc.com/saturn_he.pdf
>
> These sorts of boxes might be more compelling to more people if they
> included at least the option for HTX slots, PCI-Express slots, and/or
> on-motherboard Inifiniband support.  Or perhaps even onboard
> Intel-based gigabit on which they might choose to support GAMMA.  (Oh,
> I see the S2891 does have one PCI-Express slot, but the S2881 does
> not.)
>
> -- 
> Andrew Piskorski <atp at piskorski.com>
> http://www.piskorski.com/
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf
>