Thread View: ieee.general
1 messages
1 total messages
Started by Glenn Stafford
Tue, 10 Mar 1992 01:24
/usr/group/chicago meeting - 3/18
Author: Glenn Stafford
Date: Tue, 10 Mar 1992 01:24
Date: Tue, 10 Mar 1992 01:24
100 lines
5742 bytes
5742 bytes
/usr/group/chicago March General Meeting Announcement March 18, 1992 7:00 p.m. Sheraton Inn Naperville 1801 N. Naper Boulevard (Naperville Road) Naperville, Illinois In March, /usr/group/chicago will have a special privilege in welcoming Dr. Bruce Nelson, the Director of Technology at Auspex Systems, in Santa Clara, California. His will speak to us on: An Assymetric Multiprocessor Architecture for Extremely High-Performance Systems (or, How To Make RISC Workstations Run Faster by Removing Bottlenecks) Following is an abstract of the talk: The power of existing NFS file servers has not kept pace with increasing workstation performance because it is much easier to shrink a RISC microprocessor die than it is to spin up a magnetic disk platter. As a direct result, many users are unable to tap the full power of their server-choked workstations. This server shortfall is intensified by the NFS protocol which, while simple and stateless, is neither sophisticated nor optimal. Furthermore, mere MIPS alone do not yield excellent I/O capability. This makes it challenging to build exceptional-performance NFS servers, especially in an ares where FDDI may be too little, too late, and HIPPI-class bandwidths are desired for some applications. In research environments, the resulting opportunity is to define and refine new file systems -- such as Andrew and Sprite -- that use better protocols and semantics. Compatibility is not a priority. In the commercial realm, where 2,200,000 NFS clients demand service, one opportunity is to reimplement the NFS specification for performance and stability. Precise NFS compatibility is crucial until there is a clear and widely accepted migration path to the post-NFS world (e.g., slow migration to DCE and DCE-DFS). Auspex builds a network server that factors protocol, file, and storage processing out of Unix and executes these functions instead on dedicated 32-bit processors. Network processors support two Ethernets at full bandwidth and perform protocol processing through the NFS level. File processors run the raw 4.3 BSD file system unencumbered by context switching. Each storage processor controls ten fully concurrent, latency-optimized SCSI channels that run mirrored, striped, spanned and/or write-accelerated disk arrays. A spearate Sun3- or Sun4-compatible Unix host processor performs housekeeping and utility functions, but never touches (i.e., slows down) NFS requests. The Sun4 SPARC host processor provides a good balance of computing and I/O capabilities. This talk explores Auspex's explicitly asymmetric "functional multiprocessor" architecture (which bucks the trend of most workstation vendors and their symmetric servers) in detail. It investigates SCSI and IPI disk characteristics and shows, both analytically and empitically, that SCSI is better than IPI for NFS service. It discusses how Auspex benchmarks a full 10-processor, 80-GB, eight-Ethernet server configuration. It explains how we achieved our 1,300 average-mix NFS IOPS (I/O operations per second) rate. At an aggregate file-transfer bandwidth of 70 Mb/s, or 88% of eight simultaneous Ethernets, the Auspex machines have 4-6 times the throughput of conventional NFS file servers, even when enhanced with third-party disk and network accelerators. The talk concludes with a brief technical case study selected from the over 300 NetServers in production use. The study is Adobe Systems, home of Postscript. Adobe's dual NS 5000 configuration has 8 & 8 Ethernets, 56 & 45 GB of disk, 50+ diskless clients, and over 200 more DEC, Next, and Sun dataless workstations. About our speaker: Bruce Nelson has had broad software engineering experience with computer systems and networks. Between 1977 and 1982, Nelson was on the research staff of the Xerox Palo Alto Research Center (PARC), where he did pioneering work on distributed computing systems. He implemented the first and fastest remote procedure call (RPC) compilers (concepts now openly promulgated by Sun as ONC/RPC and OSF as DCE/RPC), developed and demonstrated practical optimization techniques for network protocol software, and build performance monitors for network file servers. Since PARC, Nelson has worked for both semiconductor and workstation companies, giving him a cross-spectrum understanding of computer system issues. He joined Auspex in 1989, where he is responsible for product planning, benchmark methodology, and technology forecasting. Nelson received his MS in computer science from Stanford University and his PhD (also in computer science) from Carnegie-Mellon University. Getting to the meeting: >From Chicago, take Interstate 88 west to the Naperville Road exit. At the light, turn right. The next light will be Diehl Road. Turn right again. The Sheraton will be on the right. Hope to see you there!
Thread Navigation
This is a paginated view of messages in the thread with full content displayed inline.
Messages are displayed in chronological order, with the original post highlighted in green.
Use pagination controls to navigate through all messages in large threads.
Back to All Threads