There was a recent article in IEEE Spectrum entitled The Trouble With Multicore that gives a really nice background on how multicore processors have taken hold in the computing industry. It also details many of the challenges multicore brings to both HPC and Enterprise computing industries. The article suggests that the advent of multicore processing was pretty much a gamble on the part of semiconductor manufacturers, and investigates what the future might hold for multicore ten years from now.

I wrote about this IEEE article on my blog Multicore in HPC - Where will we stand in 10 years? In doing so, I offered some additional insight into the future of multicore in HPC. I’m not going to rehash those ideas here, so go on over to blakegonzales.com if you are interested.

Additionally, and more importantly, I posed the same question to the High Performance Computing group on LinkedIn and there has been lots of great insight posted there as well. So much so, that I thought it would be very beneficial to share some very keen insight others in the industry have shared:


Damien Hocking: “… One thing that many articles miss is that in spite of being difficult, parallelism is really, really useful, which will in turn drive the development of skills and technology. It will force people to decompose their problems differently. Another thing that many articles don't point out is that good software engineering and programming is difficult, period. Multicore chips haven't suddenly made that hard when it was easy before, and we've had multiple-CPU motherboards for years.”

Blake Gonzales: “In so many cases, we see that parallel computations for solutions are the preferred method. We don't always hear a lot about the positives of multicore in the trade rags, just that parallel programming is a hard paradigm shift to swallow… In HPC we love parallel is so many ways.”

Avi Kcholi: “I think the two main deficiencies at present are the lack of a proper parallel operating system, which I guess is something that will come about eventually… The other problem is the frameworks that make programmers dumb. Using such frameworks programmers don’t have to think about parallel processing but simply use constructs and functions that parallel everything whether its needed or not. Having said that, multicore CPUs are here to stay and eventually will even do the trick.”

Nicolas Morey-Chaisemartin: “… In my opinion, multicore will have to change drastically to scale in next few years… Soon, we will have CPU with hundreds of cores but with a complex memory architecture! … SMP (even with NUMA) is pretty much the only thing that keeps multicore programming "accessible.” Once we lose this, we will face a real new challenge.”

Tom Mitchell: “… Multicore will solve a number of things at the operating system level once they become common. Networking and network latency along with other system IO can improve. Power consumption can improve by idling CPU cores. Additional power improvements will arrive as first and second level cache gets placed under software control.”

John Gustafson: “Folks, processors are dirt cheap! That's not what's limiting your speed. Overprovision them, and don't panic when doubling the number of processors only increases speed by 40%, especially if it only increases power consumption by 5%. Idling processors may be the next version of multithreading; you can stall on overheating just like you can stall on memory latency. The precious resource in computers is memory bandwidth (and low latency through locality). Worrying about keeping all the processors busy is like buying a piano and then feeling like it's a terrible waste not to press all 88 keys, all the time.”

Damien Hocking: “Granted, there will always be great, enormous problems that require weeks or months to run even with the best tuning and balancing. We had those problems when we had 5MHz processors too. We need those problems, they continually push the boundary of what we can achieve. BUT: The majority of the benefit from HPC and multicore is going to be delivered at the small to medium scale, a few tens to a few hundred cores.”

See what I mean? This is a really good discussion about where we are headed with multicore in HPC. My thanks to those who have contributed so far. Head on over to the discussion and let us know what you think about multicore!

-- Blake Gonzales