On the Quality of Wall Time Estimates for Resource Allocation Prediction

Today's HPC systems experience steadily increasing problems with the storage I/O bottleneck. At the same time, new storage technologies are emerging in the compute nodes of HPC systems. There are many ideas and approaches how compute-node local storage can be made usable for HPC systems. One consideration is to copy job data to the compute-node local disks in advance. To accomplish this, the allocated nodes must be known in advance. In this paper, we look at the node allocation behavior of a HPC batch scheduling system. Our goal is to determine whether it is possible to stage data in advance, based on scheduler predictions. We show that wall time estimates must be excellent to reliably predict node allocations. In reality, the required accuracy enabling advance data staging is hard to achieve. Therefore, the behavior of (standard) batch scheduler have to be modified in order to enable efficient advance data staging.