50

100

150

200

250

300

350

400

450

500

4 16 64 256

Idle time (seconds) Cache size (MB)

4K cache blocks

8K cache blocks

Figure 8: Idle time while running two instances of venus with varying cache sizes.

Execution time would be 761 seconds if there were no idle time.

7 Conclusions

While much attention has been given to CPU performance in supercomputers, the I/O system, which includes the file cache, SSD, disks, and tape storage, will play an increasingly larger role in utilizing the CPU efficiently. We have examined several high-I/O demand supercomputer applications and shown that they are highly sequential and very regular in their access patterns. This information can be used to better design a supercomputer I/O system to fully utilize a supercomputer CPU, as our buffering simulations show. With a large SSD, only one or two processes per processor are needed to keep the CPU fully utilized. While main memory sizes may not scale as fast as processor speed, SSD sizes may scale more closely as constraints on physical size, distance from the CPU, and access speed of short accesses are not as stringent for an SSD. By implementing read-ahead and writebehind in a supercomputer's file system and using a solid-state disk, a few very large processes staging data to and from secondary storage can keep supercomputer CPUs busy.

References

[1] C. Bonifas. Searching for a Unix mass storage system for a supercomputer. In Digest of Papers, pages 129{133. Tenth IEEE Symposium on Mass Storage Systems, May 1990.

[2] R. L. Henderson and A. Poston. MSS II and RASH: A mainframe UNIX based mass storage system with a rapid access storage hierarchy file management system. In USENIX | Winter '89, pages 65{84, 1989.

[3] J. L. Hennessy and D. A. Patterson. Computer Architecture: A Quantitative Approach. Morgan Kaufmann Publishers, Inc., 1990.

[4] M. D. Hill. Aspects of Cache Memory and Instruction Buffer Performance. PhD thesis, University of California at Berkeley, November 1987.

[5] NAS Systems Division, NASA Ames Research Center, Moffett Field, CA. NAS User Guide, January 1990.

[6] M. N. Nelson, B. B. Welch, and J. K. Ousterhout. Caching in the Sprite network file system. ACM Transactions on Computer Systems, 6(1):134{154, February 1988.

[7] J. K. Ousterhout, H. Da Costa, D. Harrison, J. A. Kunze, M. Kupfer, and J. G. Thompson. A tracedriven analysis of the UNIX 4.2 BSD file system. In Proceedings of the 10th ACM Symposium on Operating System Principles, pages 15{24, December 1985.

[8] V. L. Peterson, J. Kim, T. L. Holst, G. S. Deiwert, D. M. Cooper, A. B. Watson, and F. R. Bailey. Supercomputer requirements for selected disciplines important to aerospace. Proceedings of the IEEE, 77(7):1038{1054, July 1989.

[9] A. D. Samples. Mache: No-loss trace compaction. Technical Report UCB/CSD 88/446, University of California at Berkeley, September 1988.

[10] A. J. Smith. Disk cache{miss ratio analysis and design considerations. ACM Transactions on Computer Systems, 3(3):161{203, August 1985.

[11] E. Williams, C. T. Myers, and R. Koskela. The characterization of two scientific workloads using the CRAY X-MP performance monitor. In Proceedings of Supercomputing '90, pages 142{152, 1990.