• The system balance is the ratio of these two B = NPn

There is a balance point in the design of a high performance computer (HPC) system at which the contribution to performance of the mass storage (I/O) infrastructure is neither too large nor too small. That balance point has commonly been identified heuristically. Our contribution quantifies the balance point by examining the relative costs and impact of compute resources versus I/O resources. If storage is a bottleneck, improving the I/O capability can raise system utilization, increasing throughput. The balance point is where the cost of increasing throughput by adding I/O capability is the same as the cost of doing so by adding nodes. Heuristics for System Balance Figure 1a) Heuristics b) 1GB/s per TB Memory capacity is the key determinant of necessary I/O bandwidth and capacity [2]. Figure 1(a) presents the heuristic that the file system should be able to move all of memory in about 1000 seconds. Figure 1(b) presents this heuristic as applied to several HPC systems. A Model for System Balance Simplifying assumptions: • The cost model for compute and I/O capability is linear. • A job is either doing computation or I/O, and its run time is the sum of its compute time and its I/O time. • A job carries out its I/O at the optimum rate for the storage system. Definitions: • Pn is the cost of a node • Pr is the cost of a “unit ” of bandwidth • The cost of the system is nPn + rPr • U is the system utilization- the fraction of node-seconds spent in computation Results: • The marginal cost of adding compute capability by adding nodes is Pn/U • The marginal cost of adding compute capability by adding I/O capability is P