Asynchronous IO in the UM

Good meeting/seminar this afternoon on the new UM IO server, a feature that ought to massively improve the performance of STASH for model configurations with high CPU counts. The implementation details were fascinating. I particularly liked the idea of being able to dynamically balance the IO load across the servers — something that Unicos/MK never managed to do! I also thought it was interesting that using asynchronous sends to pass metadata requests to the IO servers didn't improve performance, but instead caused some collective operations to hang waiting on locks.

Fascinating stuff.