Performance of the CM-5 scalable file system

Reviewer: Mihaela Carstea

The enormous increases in the computational performance of recently delivered massively parallel systems have exposed input/output as the major performance bottleneck for many applications. The goal of this paper is to measure the I/O performance of the CM-5 Scalable File System (SFS) and to supply the low-level performance data needed to support a more extensive application characterization effort. For this purpose, the file system interfaces provided by the data parallel CM Fortran and the message-passing CMMD programming models were used. The paper is structured into six sections. The first is introductory. The second section is a brief description of the CM-5 and its parallel file system interfaces. The next two sections present benchmark methodology and analysis. To access I/O performance for the common case, a set of synthetic input/output benchmarks were developed, and SFS performance was measured under a variety of loads. The CM-5 used had the following configuration: 512 computation nodes, 15 disk storage nodes, five partition managers, and two I/O control processors. The CM-5 SDA was configured with two file partitions and had a total of 120 disks. To quantify the performance of the most salient aspects of the CM-5's SDA and SFS, three sets of synthetic input/output benchmarks were used. The benchmark analyses were grouped in three categories: CM Fortran, including file writes and file reads; CMMD, including data rates for synchronous broadcast reads, synchronous sequential writes, and hardware-dependent read and write operations; and the effects of program interactions and resource contention. The results are accompanied by illustrative diagrams. The fifth section is a discussion, including a comparison between the results obtained by the authors and those reported in other papers listed in the bibliography. For a better assessment of the results of parallel I/O tests, the authors also include a parallel systems comparison between CM-5 and another notable large parallel system, the Intel Touchstone Delta. Finally, the sixth section concludes with a summary and suggestions for future research. The paper is for readers interested in obtaining better results for parallel systems. Only with knowledge of system behavior and limitations, obtained with such synthetic benchmarks, can one reliably interpret the more complex patterns in realistic codes. The content of this work is clearly presented, and easy to understand for people with moderate knowledge of the field. A rich bibliography (15 titles) enables the reader to find other points of view about this subject.