[Bioclusters] Fwd: pre-configured clusters

[Jeffrey B. Layton]

Robert Myers wrote:

> Jeffrey B. Layton wrote:
>
>> Robert Myers wrote:
>>
>>> "The SGI Altix 350 is priced starting at $12,199. A 4-processor 
>>> configuration carries at a list price of $21,599, or $5,400 per 
>>> processor. "
>>
>>
>> I will agree/disagree. It depends upon your application(s). Some will
>> run faster on the Altix, some won't. Test your application on both
>> and then do the price/performance.
>>
>> Personally, if you need a single-system image machine and you know
>> you are not going to grow the machine, then the 350 is a great deal.
>> Otherwise, look at clusters. I think the Xserve deserves a second look
>> as does the IBM 970 (same thing really). Also look at the Xeon
>> nodes and the Opteron nodes. I think the Opterons are especially
>> promising.
>
>
> The directly comparable Xserver 455 from IBM lists for at least twice 
> as much per processor, not counting added cost if you want 
> infiniband-type performance for more than one box.  That is not to say 
> that Xeon or Opteron boxes in the Xserver series might not be a better 
> buy for some applications.  If you're consdering IBM, be sure to have 
> a quote on a 350 to show to the IBM salesman.  At a minimum, you will 
> probably get a better price. 


Good point. I've heard not good things about IBM's Opteron motherboard
(I didn't want to use the word bad). Plus I think they are putting 
gold-plated
pricing on them. A better deal is from Racksaver. I've had very good prices
from them on just about any system. (just a small, uhmm, recommendation).

>>> The Apple cluster comes standard with gigabit ethernet: an 
>>> acceptable interconnect only for a limited number of problems. The 
>>> SGI uses proprietary Numa-Link, which is at least competitive with 
>>> Infiniband, a much faster and much more expensive solution than 
>>> gigabit ethernet, but the Altix 350 looks like a really good buy, 
>>> especially if you're looking for a pre-packaged solution.
>>
>>
>> I don't see anything wrong with GigE at all. Again, it depends upon
>> your application. If you need a highbandwidth low latency interconnect
>> then something like IB, Myrinet, Quadrics, or Dolphin are the way to go.
>> If you need single system image, then NUMAflex is the way to go
>> (just a small correction - the SGI website and literatures refers to 
>> it as
>> NUMAflex, not NUMA-Link). 
>
>
> As you can tell, the amyloid plaque is already showing its effects, so 
> there is not a moment to lose. 


I think my plaque is OK, but the hair on top is just about gone. :)

>> Believe it or not, my main application
>> scales very well with Fast Ethernet (over 90% scaling for at least 300
>> processors). It's not a Bio application, but some bio apps aren't even
>> really parallel apps at all - just HUGE amounts of serial processing 
>> with
>> varying data sets. 
>
>
> Agreed that not everything needs a low latency, high-bandwidth 
> interconnect.  For those who do, and who want a pre-packaged solution, 
> the Altix 350 just happens to be what appears to be an exceptional buy 
> at the moment, as compared to other prepackaged low-latency, 
> high-bandwidth solutions currently on the market.


Agreed. Good point.

>> Oh, another thing about NUMAflex, it's not the same across their
>> product line. While I can't confirm this, I was told that the Altix
>> 350 had a ring type connection for NUMAflex. The larger and more
>> expensive machines had a fat-tree type connection for NUMAflex.
>> If you're going this route, be sure to ask (and be sure to test!). 
>
>
> Altix 3700 has a very expensive two-plane, fat tree switched NUMAflex 
> network.  The Altix 350, which scales up to 16 processors under a 
> single system image, uses a ring NUMAflex network as does the Altix 
> 3300.  From usenet posts made by someone associated with SGI, I infer 
> that the boxes have pairs of Itania sitting on a standard Intel 
> processor bus sharing a memory controller.  An eight processor 
> configuration involves at most three NUMAflex hops, from the ASCII-art 
> diagram he posted.  No one has ever said so officially, but I am sure 
> you can cluster the single-image clusters, but to do that you have to 
> start buying stuff like router bricks.  The 350 is likely to be most 
> attractive to people who want relatively uniform access to large 
> datasets in memory with a relatively modest number of processors. 


This jives with what I've heard from some folks who work in
California (I won't name my company yet, but some people
already know). These people were concerned about the number
of hops in the NUMAflex boxes because their code has been
parallelized yet (I'm working on it if they will give me the
budget to finish it and test it properly).

Have a good day!

Jeff