Characterizing the Balance of Parallel 1/0 Systems

Abstract

High pzrformance 110 subsystems are a key element in parallel computer systems that intend to compete with traditional supercomputers in the solution of large scientific and engineering problems. The hardware and software organization and perfomiance of these U0 subsystems are fundamental issues in parallel computer systems that have not been explored in depth. Two commercially available parallel file systems are the Intel iPSC/2 Concurrent File System (CFS) and the NCUBE/ten NChannel board and disk farm. Both systcms are aimed at support of high volume, large block I/O of the type typical of large scientific computations. The evaluation of these systems has proved difficult. There are many parameters affecting performance and the system dynamics are quite complex. In this paper we examine a method of quantifying the balance of an 1/0 system, that is, how well it services I/O rcquests with respect to fairness and distribution of overheads. One may gauge the degree of balance in a systcrn by asking: When resources become saturated, is the bottleneck felt equally by each process or are some processes given preferential service? This paper explores a simple yardstick of system balance. 1. Quantifying and Measuring Parallel I/O Suppose that we have p processes reading (writing) a file of N bytes in parallel. Each process i reads (writes) N n bytes in time ti where n = -. The individual data P transfer rate of a particular processor i is given by ri = z. The average individual data transfer rate is ti given by 7 = 1 firi. P I = There are at least two reasonable measures of the aggregate data transfer rate of the p processors. In the first case, we sum the data rates of the individual processors. This gives rise to the quantity ri called the maximum sustained aggregate rate (ma-SAR). We call this I = 4 tThis research was supported in part by JPL Contract #957721 and by the Department of Energy under Grant DE-FG05-88ER25063. measure the “maximum” rate because, by construction, it assumes that each processor i contributes a rate ri and all processors contribute during the same time instant, however brief. From the definition of F above, we see that max-SAR = ri = p 7 . (1) I = f i This interpretation is illustrated in Figure l(a). In the second case, we consider that all N bytes move through the system in z = max ti time units. That is, the entire file is not transferred until the slowest processor finishes reading (writing) its partition of the file. This gives rise to the quantity called the minimum sustained aggregate rate (min-SAR). We call this a “minimum” rate since this is the rate that an outside observer will perceive as the rate at which the entire processor ensemble is operating. From the definitions above, we see that 1