Parallella, Raspberry Pi, FPGA & All That Stuff

Good to see someone willing to take the plunge. I thought of simply making a cluster of Raspberries. Costs got out of hand. There is so much on the card not needed for a cluster such as video outs that they cost more than they need to the cables to match the fixed connectors become a significant cost. Seems to me the Raspberry folks could put a mess of chips on a larger card and have a cluster to sell rather easily.

But what has my caution is the assumed ease of programming. The issue of parallel processing has never been the language but how to make use of parallelism. Optimally a quad core computer should be four times faster minus 10% for overhead, 360% faster.

A simple example is POVRay which was ported to multithreaded a couple years ago. One a quad core the results are about 250% faster. However alternate software could divide resulting image into four quadrants, assign one per core and get 360% and then put it back together. Simply adding multithreading is much simpler and errors can be blamed on the chip. As I understand it dividing the image is what the CUDA cards are doing. And this is what one BOINC work unit per core is doing. It isn't parallel processing. It is being treated like for computers on one chip. It is also without all the superfluous stuff that comes with a Raspberry.

Which sort of says that if the problem being addressed is not naturally parallel like rendering an image it isn't really worth the effort to do more than multithreading. The language used hardly matters.

While multicores are great for BOINC projects at one WU per core for most computer users they cannot tell the difference between two and more than two cores there is so little extra improvement. Workload is like 60/30/10 with the fourth core for the OS if stuck with Windows.

So the best use I see for a cluster like this is for programmers to find a way to program them in a general way such that everything can get the 360% instead of the chip controlled 250%.