Friday 4 November 2011

My SC11 diary 10

It seems I have been blogging about SC11 for a long time - but it has only been two weeks since the first SC11 diary post, and this is only the 10th SC11 diary entry. However, this will also be the final SC11 diary blog post.

I will write again before SC11 in HPC Wire (to be published around or just before the start of SC11).

And, then maybe a SC11 related blog post after SC11 has all finished.

So, what thoughts for the final pre-SC11 diary then? I'm sure you have noticed that the pre-show press coverage has started in volume now. Perhaps my preview of the SC11 battleground, what to look out for, what might emerge, ...

Let's start with processors. GPUs are no longer the new kid on the block - they have been hyped and evaluated long enough to be considered part of the HPC landscape now. ARM and DSPs have replaced GPUs as the emerging and potentially disruptive HPC technology to find out about, with promises of low power, better FLOPS/Watt, cheaper HPC, etc.

Even Intel is getting some attention in the "new" architectures arena with MIC. Intel's manycore product family (Knights Corner) is nearing availability, and has already won some deals (e.g. Stampede at TACC). NAG's own experience so far in the MIC early evaluation program is that the MIC is relatively easy to port code to.

And, of course, the traditional CPU teams, Intel Xeon (e.g. Sandy Bridge) and AMD Opteron (e.g, Interlagos) will be making their case as the HPC processors to choose. AMD's Bulldozer/Interlagos is sufficiently different from previous processors (shared FP unit, FMA4, etc) to be thought of as "new".

We have CPUx86 vs MIC vs GPU vs ARM vs DSP. So, the processor/architecture battleground looks vibrant and interesting. Let's see who emerges as the popular candidate after SC11.

What about software? Well, almost all of those architectures will require developments to your software in order to effectively exploit the new processor technology. For MIC, the porting looks OK, but vectorization is needed for good performance. For GPUs, it is not just a case of re-writing in CUDA or OpenCL, but of reconsidering the choice of algorithms used to extract much more parallelism at the node level. If that works, then performance can be very good. For ARM - who knows yet? But software (portability, performance, ease-of-performance, etc.) will be the key, no doubt.

Almost independent of your choice of processor, scale will still be an issue. Whether hundreds of GPU accelerated nodes or thousands of pure CPU nodes, software will have to scale to get the performance. Tools and expertise to support this will be critical. What software tools will emerge from SC11 as important players in the development of scalable software - from desktop to multi-petascale? Well Allinea is one example that has been making some pre-show noise, but there are others. Also worth keeping an eye out for news on the expertise - programs to develop and encourage careers in HPC, training opportunities, recruitment, etc. [Commercial message: NAG can help with expertise - provide parallel programmers to scale your applications, training, consulting, ...]

Big Data. HPC Analytics. SC11's theme is related to big data and HPC - and the pre-show noise seems to have produced plenty of content on big data and analytics using supercomputers. In my cynical days, I simply say this is merely the new buzz. Cloud is too blah now (been on the marketing agenda for too  long). Same for Green. Both are still very relevant to HPC, but they aren't "fresh". So along comes HPC Analytics or Big Data. Massive and often volatile data sets generated by or processed by supercomputers is both a key challenge and an opportunity. One well worth looking out for solutions to at SC11. Products and solutions for large data will be a big part of SC11 exhibition and discussions. As will the hardware and software for processing massive data. R&D in this area will be important too. But I fear there will be some over-use of the "big data" topic, as with cloud and green, which will only work to obscure the real problems and potential.

Rumours. The fun bit of SC. What news or gossip will break during SC11? Key people moving between companies. Large deals announced. The replacement for IBM Blue Waters? The next big non-USA supercomputer. A new processor? Acquisitions? Sometimes the Top500 provides the big news and the surprises. I have heard whispers of some some things that may emerge, and there is always scope for unexpected announcements, leaks, or gossip.

What else? The "missing middle". Broadening HPC to organizations and users outside the big labs, supercomputer centers and large companies. This important issue might make headway at SC11. I will be involved in several discussions around this topic (including presenting on the Intel booth). And, let's not forget Exascale - I'm sure that will get plenty of attention at SC11 too.

What are you looking out for? What do you think will emerge as winners and losers?

Finally, if you can, make sure you catch the HPC Analyst Crossfire session Friday - I was a panel member at ISC11 and it was good fun - but more importantly, Addison Snell has pulled together an excellent panel for SC11 and it will be well worth your time.

I'll have a busy SC11 but if you want to find me to hear more about my opinions on the HPC world, berate me over a blog post, or learn more about NAG's HPC services & consulting, then please do get in touch.

See you in Seattle.

No comments: