Actually a problem few people are aware of and are working on is the problem of IO. While there is the old quip of supercomputers being devices to turn CPU boun

Actually a problem few people are aware of and are working on is the problem of IO. While there is the old quip of supercomputers being devices to turn CPU bound problems into IO bound problems, it is a real problem to keep a well tuned HPC application on a supercomputer fed with data. And even if you are lucky and have an application with with a small input dataset, it gets hard to write snapshots (e.g. for checkpointing) once you go beyond 100000 CPUs or so. And at that point you really want checkpointing, because the failure probabilty of the nodes rears it's ugly head.