The heat produced by modern CPUs means that chip designers can no longer crank up the clock speed. But they can fit more processing cores on a chip. 2-core and 4-core x86 chips are already standard; Sun's SPARC chips have 8 or 16 cores, and Intel recently announced an 80-core prototype.
To get the best performance out of multi-core chips, especially once you get beyond 2 or 4 cores, you either have to write applications to make best use of those cores or write compilers and dynamic optimising systems that automatically transform your code to get that performance. The problem is, most programmers aren't very good at this and most of the tools they use aren't brilliant either.
I think this has immediate policy implications. CS departments should immediately, if they haven't already done so, make the teaching of parallel and distributed programming a required component of their undergraduate courses. Companies and professional bodies should encourage their staff to retrain. For this retraining, someone has to prepare and give the courses. Some government funding might kickstart the process and so ensure that the UK doesn't fall behind.
Several pundits, including Tony Hey of Microsoft, have suggested that experience from the supercomputer world could help, because supercomputers have been massively parallel for years. There is some truth in this, and organisations such as EPCC are already teaching parallel programming techniques. But there are differences too. The economics of the supercomputer world are rather old fashioned; there, the computer is still the expensive part of the system while programmers' time is relatively cheap. So programmers often tune each application for each new machine.
You can see this philosophy in much of the scientific grid world: people submit jobs that require a certain number of processors to run. Think about that. They're not asking for a certain amount of processing power, nor for their jobs to run in a certain time or for a certain cost. Their applications are programmed to run on a specific number of processors. To anyone outside that community, this approach is clearly crazy, and it certainly won't transfer to a world where multi-core processors are plentiful and programmers' time is expensive.
So the distributed (grid) world and the multicore CPU world face similar problems. Both need new programming models and tools. Fortunately there are many computer scientists who are investigating better ways to program distributed and parallel machines. I can't attempt to represent the field but some of my friends and colleagues are working on such problems. For example, Murray Cole at the University of Edinburgh has developed a model for transforming parallel programs on to different numbers of processors. In 2002, a team from Microsoft Research in Cambridge incorporated modern concurrency abstractions into a research version of C# - replacing the usual locks, semaphores and critical regions that Tony Hoare invented 40 year ago.
Here at the e-Science Institute, we are about to launch a theme - a series of workshops and visitors - on Distributed Programming Abstractions. I certainly hope that this will address some of the questions raised above and that soome of the people reading this will contribute. Another useful resource is this wiki at Berkeley, which also includes an interesting white paper.
To get the best performance out of multi-core chips, especially once you get beyond 2 or 4 cores, you either have to write applications to make best use of those cores or write compilers and dynamic optimising systems that automatically transform your code to get that performance. The problem is, most programmers aren't very good at this and most of the tools they use aren't brilliant either.
I think this has immediate policy implications. CS departments should immediately, if they haven't already done so, make the teaching of parallel and distributed programming a required component of their undergraduate courses. Companies and professional bodies should encourage their staff to retrain. For this retraining, someone has to prepare and give the courses. Some government funding might kickstart the process and so ensure that the UK doesn't fall behind.
Several pundits, including Tony Hey of Microsoft, have suggested that experience from the supercomputer world could help, because supercomputers have been massively parallel for years. There is some truth in this, and organisations such as EPCC are already teaching parallel programming techniques. But there are differences too. The economics of the supercomputer world are rather old fashioned; there, the computer is still the expensive part of the system while programmers' time is relatively cheap. So programmers often tune each application for each new machine.
You can see this philosophy in much of the scientific grid world: people submit jobs that require a certain number of processors to run. Think about that. They're not asking for a certain amount of processing power, nor for their jobs to run in a certain time or for a certain cost. Their applications are programmed to run on a specific number of processors. To anyone outside that community, this approach is clearly crazy, and it certainly won't transfer to a world where multi-core processors are plentiful and programmers' time is expensive.
So the distributed (grid) world and the multicore CPU world face similar problems. Both need new programming models and tools. Fortunately there are many computer scientists who are investigating better ways to program distributed and parallel machines. I can't attempt to represent the field but some of my friends and colleagues are working on such problems. For example, Murray Cole at the University of Edinburgh has developed a model for transforming parallel programs on to different numbers of processors. In 2002, a team from Microsoft Research in Cambridge incorporated modern concurrency abstractions into a research version of C# - replacing the usual locks, semaphores and critical regions that Tony Hoare invented 40 year ago.
Here at the e-Science Institute, we are about to launch a theme - a series of workshops and visitors - on Distributed Programming Abstractions. I certainly hope that this will address some of the questions raised above and that soome of the people reading this will contribute. Another useful resource is this wiki at Berkeley, which also includes an interesting white paper.
Comments
Thoams Sterling addresses these issues too. His recent keynote points out some of the differences as well as the similarities. A major difference is the bottleneck in memory access on multi-core CPUs. See http://www.cct.lsu.edu/~tron/ICCC06EndNoteFinal.pdf.