Southampton engineers a Raspberry Pi Supercomputer

Posted Sep 14, 2012 20:53 UTC (Fri) by daglwn (guest, #65432)
In reply to: Southampton engineers a Raspberry Pi Supercomputer by Cyberax
Parent article: Southampton engineers a Raspberry Pi Supercomputer

In clusters, maybe. Not in a real supercomputer. There is a LOT of communication that happens on a big IBM or Cray machine. That's why those companies build custom interconnects.

Infiniband is _maybe_ somewhat adequate. For certain codes. And certain problem sizes.

Southampton engineers a Raspberry Pi Supercomputer

Posted Sep 14, 2012 21:31 UTC (Fri) by daglwn (guest, #65432) [Link]

100G Ethernet is 100 Gbits/s. With terrible latency and horrendous congestion issues.

On a supercomputer we're talking bandwidths of ~10 GbBytes per port (meaning ~150 G-bytes into/out of a node, compared to 100 G-bits/s for 100G Ethernet), scaled up to a bisection bandwidth of ~5 T-Bytes/s with latency in the 1us range compared to ~3ms for 100G Ethernet.

And then there's I/O bisection bandwidth of 1 TByte/s or more.

Numbers from here:

http://www.theregister.co.uk/2010/05/25/cray_xe6_baker_ge...
http://www.nccs.gov/jaguar/
http://www.networkworld.com/news/2012/091212-brocade-100g...

Note that Jaguar is a relatively old machine these days so the HPC numbers are conservative.

It's mind-boggling just how huge these machines are, in FLOPS, I/O and network capability.

Southampton engineers a Raspberry Pi Supercomputer

Posted Sep 14, 2012 21:47 UTC (Fri) by dlang (guest, #313) [Link] (2 responses)

the problem boils down to how you define the term 'supercomputer'

There was a time when a system with the specs of this system would have been a supercomputer with no arguments from anyone.

you can try to define a supercomputer by it's technical specs (including interconnect speed, etc), but you have to constantly redefine your spec list.

or you can define it based on the architecture and how it's used.

While the performance of this system and the systems in the 'top 100' list are wildly different, they are actually far closer to each other in terms of architecture and how they are used than any of them have in common with the machine you are using to view this post, even if that machine is superior if you just look at the specs (flops, storage, memory)

Southampton engineers a Raspberry Pi Supercomputer

Posted Sep 14, 2012 23:31 UTC (Fri) by daglwn (guest, #65432) [Link] (1 responses)

> the problem boils down to how you define the term 'supercomputer'

Here's one attempt: A machine the can run the large challenge problems of the day in a reasonable amount of time.

> they are actually far closer to each other in terms of architecture and
> how they are used

If by, "how they are used" you mean, "programmed with MPI," sure. I can program the multiple cores on my desktop with MPI too. But the architecture of a Jaguar-class system and that of this Raspberry Pi cluster are vastly different. Sure, they have processors, memory, network and I/O but so does every other computer in existence.

Think of it this way. This Raspberry Pi thing is about taking some existing components and hooking them up in an interesting way. Building HPC systems is about designing a SYSTEM from the ground up, tailored to meet the most demanding computing challenges of the day.

Southampton engineers a Raspberry Pi Supercomputer

Posted Sep 14, 2012 23:42 UTC (Fri) by dlang (guest, #313) [Link]

umm, many supercomputers were built by taking existing components and hooking them up in much the same way that this thing was. special hardware is optional, not an inherent requirement.

if a system is only a supercomputer if it can handle the most demanding challenges of the day, is something 1/10 the power of the most powerful device still a supercomputer?, what about 1/100 the power?

note that the "top 500 supercomputer" list has the top entry with a rating of 16324 and the bottom entry with a rating of 60. how can a machine so weak still be considered a "supercomputer" according to your criteria

the Raspberry Pi thing being discussed here is a cluster of machines, not a single machine. It has separate memory, storage, network, etc for each node. This is a vastly different environment to work with than a desktop system.

No, this cluster isn't going to do any groundbreaking research or solve any "large challenge" problems. But as something to teach people about supercomputers and let the experiment with and learn what does and doesn't work for HPC computing, this is a much better thing to use than an equivalently priced SMP system.