Does anyone have some inherently serial problems coded in SciDB they would be willing to share?
Algorithms that can be parallelized don’t worry me. At some level, throw enough hardware at it, and it will go fast. (Oversimplification, I know, but not as worrisome as my next point…)
What worries me more is algorithms that have to run serially. Then suddenly, the language architecture really matters. For example, Exadata can throw many CPUs at anything you run in Oracle. But if it’s inherently serial, and scanning the table, you’re in trouble.
Conversely, one could argue, that a good architecture makes all the difference, such as how SciDB lays out its data on disk, uses chunks and overlaps, and has other constructs that let it perform very fast–even when the algorithm in question is inherently serial. But I’d want to run some actual code to prove it.
What I’m going for hear is a test base of inherently serial algorithms that I can run on SciDB and on other solutions to see head-to-head how they perform.
All assistance appreciated!