Supercomputers are comprising more and more processors and these processors are increasingly heterogeneous, with differing performance characteristics. The conventional programming models assume that all nodes run in lockstep. Thus, applications run at the speed of the least powerful processor. We introduce DesyncFS, a new programming model based on the block abstraction of traditional file systems. It virtualizes the performance characteristics of processors; this allows their heterogeneity to be hidden. We show that DesyncFS allows cluster throughput to scale with average processor throughput instead of being limited by the slowest processor.
| 