A mass storage system for supercomputers based on Unix

Abstract

This paper presents the design, implementation and utilization of a large Mass Storage Subsystem (MSS) at the Numerical Aerodynamics Simulation (NAS) at NASA-Ames Research Center. The MSS supports a large networked, multi-vendor Unix-based supercomputing facility.The MSS at Ames Research Center provides all processors on the Numerical Aerodynamics System Processing Network (NPSN), from workstations to supercomputers, the ability to store large amounts of data in a highly accessible, long-term repository. The MSS uses Unix System V and is capable of storing hundreds of thousands of files ranging from a few bytes to 2 Gigabytes in size. This utilization of Unix, to mask the underlying MVS system, reduces the training requirements for the end user and allows the use of existing Unix networking solutions such as RCP, REMSH, and TCP/IP with locally developed performance enhancements. The systems are currently networked via the NSC HYPERchannel and Ethernet. To assure data integrity and to accommodate spanned volume files the IBM MVS Operating System was implemented so that it is transparent to the Unix user.Since the MSS was released for production at NAS in June of 1987, experience with the system has led to design improvements and enhancements which have resulted in more than an order of magnitude in performance improvement. The current production system is capable of data transfer rates in excess of 6,900 Kilobits per second end-to-end from the Cray 2 supercomputer disk to the MSS DASD. Future improvements include parallel data transmission across multiple NSC HYPERchannel adapters and data-striping on the Amdahl 6380 double density disk drives. Initial performance tests have demonstrated a further increase to 17,800 Kilobits per second when these improvements are placed into production.This article is based upon a less extensive paper presented at the IEEE Symposium on Mass Storage Systems, October 1988 [1].