Robust I/O Performance in River

Remzi H. Arpaci-Dusseau and Eric Anderson


The key problem with cluster applications that perform parallel I/O is that they are very sensitive: in a quiescent system, performance is excellent, but a small perturbation on a single workstation leads to a large-scale performance hit. We would like to move applications away from this meta-stable performance point into a more stable regime, where small perturbations lead to small (or no) performance degradation. To assist applications in this direction, we propose a system known as "River". By defining a higher level, more flexible interface, and a dynamic, environment-aware system underneath, we plan to provide robust parallel I/O to a range of interesting applications (decision support, scientific, etc.). The River mantra is "move the I/O to the computation", and draws on work from both the parallel I/O and task queue literature.
Click here for the talk slides.

Click here to go back to the Finale schedule.