Going back to the HPC education issue I mentioned in my previous post, that actually touches on the other theme that was clear today at the sessions at the 23rd HPCC conference.
Seems like almost no one argues against that the future for continued increased performance in HPC is moving towards a future with more multithreading, multicore, many-core, GPUs and other accelerators, often in a heterogenous mix of thousands (or millions eventually) of each.
This is not a panacea however, there certainly are problems to be solved in those areas as well, the infamous memory wall and energy consumption would be two of them.
The biggest challenge in my mind though is on the software side. Our middleware, tools and applications are just not keeping up. We don’t have the software technology today that makes it easier to automatically take advantage of the inherent parallelism in the hardware infrastructure. We’re today edging into the Petascale era and providing essentially assembly level programming tools. That won’t work for the next level, closer to Exascale.
We need to invest in software that on one hand hides the underlying complexity and makes it easy to scale and on the other hand makes it possible to state the problem to be solved that is close to the natural representation of it. Much like Fortress allows mathematical notations to be used to easier represent equations. We need to bridge the gap between the domain knowledge that can describe the problem and the low level “magic touch” that is needed to get code to scale.
It’s not that this is new news. Many people have pointed this oput, but we don;t seem to make progress towards a solution. It’s not that we as a “collective ostrich” is hiding our head in the sand and hoping it will go away. It won’t.
The problem is that there’s no business case for a single software vendor to take on this huge challenge. This is an area that definitely requires government funding and industry wide attention.
One speaker suggested that HPC needs to be elevated to the same importance as a nationwide energy strategy. It’s that important. I tend to agree with him. We need to do whatever it takes to start to make progress in this area.
I also intend to continue twittering tomorrow under my Bearcrossings twitter id.