Ask HN: NoSQL / Distributed File System with Broadcast, Map Reduce, etc

I am looking for a distributed file system or similar on which I can build distributed image processing (for astronomy). Some distributed file / batch processing systems exist in academia, but I was wondering if there was any way to use recent NoSQL work. A typical process involves a small number of calibration files that would need to be broadcast to all nodes (these files are as large as each image); a step in which multiple files can then be processed in parallel; finally a "reduce" that combines those into a single result. For efficiency I would prefer memory mapped files (binary blobs) that can be mutated. In addition, I need to be able to write "jobs" that can call out to GPUs (so would prefer C but could use, say PyOpenCL). Any pointers appreciated (thanks).