Computing recommendations
Introduction
From time to time, people ask what they'd need to set up a computing cluster for fMRI data analysis. I used to collect my thoughts on this boring but important subject on this web page, in the hope that they would be helpful to someone. However, those thoughts tended to get outdated very quickly. Both imaging and computing are fast-moving fields. My advice is now this: if you need help with computing, talk it over with someone who understands your needs well.
The remainder of this page contains some vague musings that I didn't have the heart to remove.
Some Preferences
This section was part of the original version of this document, authored by Dan Kimberg and written in the first person. Feel free to edit it, but since it describes positions that are closely related to VoxBo's design philosophy, it would be better to add alternative viewpoints than to try to reshape this section.
That said, here are some thoughts on computing.
I don't like monolithic central computing. I do like distributed networks made up of cheap machines. I don't like depending on commercial software to do scientific research. I do like depending on free and open-source software. I think these preferences are well motivated, and I'm more than happy to defend them (try the "discussion" tab above). But if you must have a supercomputer or you must run an operating system that's "supported" by a big corporation, this document isn't for you. If you already agree with me on these points, and don't feel like a pep talk, skip to the next section. If you're not sure where you stand on these issues, let us break some of these preferences down a bit.
The most useful software for the analysis of neuroimaging data runs best under variants of UNIX (e.g., Linux, OSX, and Solaris). Most free software will run on just about any UNIX variant, and commercial software is headed in that direction. But in general, you'll meet the least resistance with Linux. Correspondingly, we prefer PC-style hardware. For practical purposes, that means machines designed to run Windows. There are few markets as competitive as the PC market, and it happens to be the home platform for Linux. Although you won't get the single fastest machine buying PC hardware, you often get the most computing muscle per dollar.
This general approach has some distinct advantages.
- Redundancy. It's unlikely for ten computers to die at the same time.
- Ease of upgrade. With a large number of cheap computers, you can upgrade your facilities more continuously. Every time you have some extra cash on hand, buy a new machine. When your slowest machine can't keep up anymore, find another use for it.
- Cheap muscle. If you want the fastest single machine available, buy a supercomputer. If you're more concerned about computational power per dollar, clusters of cheap machines are probably your best bet. This is of course only true if your software takes advantage of clustering reasonably well. VoxBo does.
- Multi-purpose. x86-based machines can serve both as desktop machines and as crunching machines, sometimes simultaneously.
VoxBo does run on OSX, and could probably be made to run on Solaris easily if we had a Solaris box. It even runs under Windows via the [Cygwin] environment.
Please note that these preferences are not motivated by some advocacy agenda. Members of team VoxBo have collectively written plenty of code for MacOS (pre-OSX), Windows, various UNIX variants, and numerous less popular operating systems (e.g., AmigaDOS). We've used them all as desktop OSes, not just for a few minutes, but regularly for years. They all have strengths and weaknesses. The only one that really holds a warm spot in our hearts (okay, my heart -dan) is AmigaDOS. We go with what works for us, and Linux works well for image analysis.
Last modified: 19:17, 25 April 2010 (UTC)
