For a wide variety of computational tasks, disk I/O continues to be a serious obstacle to high performance. The focus of the present paper is on systems that use multiple disks per SCSI bus. We measured the performance of concurrent random I/Os, and observed bus-related phenomena that impair performance. We describe these phenomena, and present a new I/O performance model that accurately predicts the average bandwidth achieved by a heavy workload of random reads from disks on a SCSI bus. This model, although relatively simple, predicts performance on several platforms to within 12% for I/O sizes in the range 16--128 KB/s. We describe a technique to improve the I/O bandwidth by 10--20% for random-access workloads that have large I/Os and high concurrency. This technique increases the percentage of disk head positioning time that is overlapped with data transfers, and increases the percentage of transfers that occur at bus bandwidth, rather than at disk-head bandwidth.

This paper discusses our implementation of Rajasekaran's (l,m)-mergesort algorithm (LMM) for sorting on parallel disks. LMM is asymptotically optimal for large problems and has the additional advantage of a low constant in its I/O complexity. Our implementation is written in C using the ViC* I/O API for parallel disk systems. We compare the performance of LMM to that of the C library function qsort on a DEC Alpha server. qsort makes a good benchmark because it is fast and performs comparatively well under demand paging. Since qsort fails when the swap disk fills up, we can only compare these algorithms on a limited range of inputs. Still, on most out-of-core problems, our implementation of LMM runs between 1.5 and 1.9 times faster than qsort, with the gap widening with increasing problem size. 1. Introduction Researchers in many fields often wish to solve problems that are too large to fit into main memory, but traditional in-core methods are generally unable to handle very large se...