As SGI’s Chief Technology Officer, I spend around half my time with our Sales Reps – meeting customers, talking about our long-term plans, and listening to the issues they have with today’s systems and their needs for new solutions. This feedback is crucial to the other half of my work – with SGI’s senior engineers, looking for innovations, improvements, breakthrough ideas, and bringing customer feedback into the company to influence our next generation of machines.

SGI has always been customer-driven, but now we are even more-so, especially with major customers who want to be more involved and feel proud that they can help to steer us. Many of our customers have similar requirements, and we also have a smaller group who – either because they’re doing more extreme work or have bigger plans – are starting to talk about systems with upwards of 100,000 or even one million cores. Nevertheless, what everyone, big or small, has to deal with is the new challenge of “many-core” processor chips, and the fact that they can no longer rely on ever-faster clock speeds to give them the performance improvements they’ve come to expect.

Meeting the challenge of many-core
For 20 years, clock speeds kept improving, and the resulting increase in power consumption was kept under control – at least to an extent – by decreasing voltages. Eventually, however, voltages decreased to the point where they couldn’t go much lower, and so the industry had to switch to a different approach, i.e. many-core.

This is bringing a whole new set of problems for the software industry. In the past when your code ran on one processor, if your needs weren’t that extreme and you wanted to process more data, you could just wait for a faster processor to come along. Soon it won’t be that way. So a big issue for the industry is to realise this and put resources into scaling software codes.

In HPC, at the extremes we have two types of applications: embarrassingly parallel such as cluster-based render-farms, which have virtually no communication and so will be fine in this new world of scaling codes; and communications-intensive like fluid dynamics or molecular dynamics applications, where as you move one molecule you affect many things because many things are interrelated.

For communications-intensive applications, if software writers solve the issue of distributing codes, the problem then shifts to us – to build hardware that can deal with the increasing communications between processors resulting from more and more cores. This is why we believe our large single memory approach will become increasingly attractive – even for cluster applications – because as applications are forced to scale to more and more cores, there will be a point where the cluster approach will no longer work, because no matter how fast the processor and cluster interconnects, they will be bogged down dealing with communication. Large single memory can alleviate that problem.

Of the different types of communications-intensive applications, there is one that even large single memory cannot deal with totally – “global collectives”. To deal with these effectively we will have our next-generation Ultraviolet platform, which is due in the second half of 2009. Ultraviolet’s goals are to lower the cost of SGI® Altix® 4700 compared to clusters; and solve the global collective communication problem by adding a hardware MPI Offload Engine (MOE) to offload these collectives from the CPU and run them in parallel.

Already our more “extreme” customers see this as the way forward, because they’ve profiled their codes and seen the worrying trend that a higher and higher percentage of CPU time is spent dealing with communications. In fact, a number of these customers have already pre-ordered Ultraviolet systems from us. Following in the footsteps of these “extreme” customers will be our mainstream customers as their own application core count increases. I don’t think the majority of the industry has fully realised the implications of this yet, and it is therefore a major opportunity for SGI

Virtualisation: Server, Storage and Visualisation
Just as many-core is transforming the challenges for the server industry, so visualisation has been flipped on its head.

In the ’80s, a typical model would have something like 100,000 triangles, and these would be put through a graphics system to produce a million pixels. Today we have much more complex models with a billion triangles, but on the display side a laptop probably still has just two million pixels. So the visualisation process has flipped from one of expansion (100,000 to a million entities) to compression (a billion to two million entities). The goal now for high-end visualisation is therefore to build systems that can bring in huge amounts of model detail, process them quickly, and send them through commodity graphics cards to produce interactive pictures for the user.

Also, now that the model is so big, you don’t want to move it. You want to keep it where the server is, bring the graphics card into the server, and then push pixels to wherever you need them through our Wide Area Visualization Environment (including Vizserver).

The industry talks about server virtualisation and consolidation, and that’s well understood. As I’ve explained above, we’re seeing the same with visualisation. And in storage, disk capacity is growing, but the speed at which you can read a disk isn’t keeping up. So even though disk capacity is growing quickly, in our increasingly digital world where people are dealing with petabytes of data, it’s becoming unaffordable to keep everything on disk.

Some customers have reached that point, but they don’t want to go to tape because of the inconvenience of retrieving data. So this is creating a new business of storage virtualisation, where disk storage is used only as a cache, and our SGI® DMF storage virtualisation software manages a mixed disk and tape environment in such a high performance, smart and seamless way that users think everything is still on disk.

Greener than green
Two other major trends are the huge performance benefits of running Oracle and MySQL databases entirely in our large single memory systems, while also leveraging our scalable I/O subsystem; and of course “Green” issues.

SGI started designing highly energy-efficient systems so we could pack as many as possible into a rack to maximise density. As a result, our power supplies are the best in the industry, and already exceed the Climate Savers 2010 requirements at full load.

Just as importantly, we also play a role in another way, by supplying systems to – and working with – many of the organisations that are running computer simulations which are alerting the world to climate change, its implications, and designing means of alleviating it. Being green is a responsible thing to do, and we are appreciative that we can play a role in helping this realisation and subsequent alleviation.

FURTHER INFORMATION
www.sgi.com