In the petascale era, the storage stack used by the extreme scale high performance computing community is fairly homogeneous across sites. On the compute edge of the stack, file system clients or IO forwarding services direct IO over an interconnect network to a relatively small set of IO nodes. These nodes forward the requests over a secondary storage network to a spindle-based parallel file system. Unfortunately, this architecture will become unviable in the exascale era. As the density growth of disks continues to outpace increases in their rotational speeds, disks are becoming increasingly cost-effective for capacity but decreasingly so for bandwidth. Fortunately, new storage media such as solid state devices are filling this gap; although not cost-effective for capacity, they are so for performance. This suggests that the storage stack at exascale will incorporate solid state storage between the compute nodes and the parallel file systems. There are three natural places into which to position this new storage layer: within the compute nodes, the IO nodes, or the parallel file system. In this paper, we argue that the IO nodes are the appropriate location for HPC workloads and show results from a prototype system that we have built accordingly. Running a pipeline of computational simulation and visualization, we show that our prototype system reduces total time to completion by up to 30%.

A new technique is proposed for analyzing models of (paged) virtual memory management This technique, which is based on recent developments in the theory of optimal control, permits the use of a very general model of program behavior.In contrast to existmg studies of program behawor, staUstically confined to Markov models, a general jump stochastic process is used here to describe the page reference generator A model of memory management is formally defmed as three component processes: the program behavior, the memory allocation, and the control process.Equations linking the evolution of the memory allocation process with the other two processes are derived.Necessary and sufficient conditions for an optimal control policy are given as a set of optunality equations.Although these equaUons can be numerically solved for small s~ze problems, an analyUc solution is presented for the case of stationary ranking of pages.The equations of motion of the memory allocation process are used m apphcations where the performance of a given control pohcy is to be assessed for specified program behavior Another use of the proposed technique is the formulation of a Markov phase behavior model The hfetime function calculated in this model is shown to be m good agreement with empirical results.

The history of modelling of the address sequences generated by computer programs (often termed “program behavior”) follows a familiar pattern: the better a hypothetical model fits experimental evidence, the less amenable it is for calculation. In this paper programs that generate successive page references that can be described by a first order Markov chain are considered. We produce a closed form expression for the distribution and usable expressions for the first moments of the steady state size of their working set of pages. These expressions are also specialized for the independent reference model and the Easton model. Only standard Markov chain theory is used.