hewlett-packard UNITED STATES
Skip site-wide navigation link group hewlett-packard home products and services support solutions how to buy
hewlett-packard logo with invent tag line - jump to hp.com home page
End of site-wide navigation link group
 
printable version
digital technical journal online
hp labs skip lorem ipsum dolor navigation menu link group
search
contact hp
introduction
foreword
table of contents
online issues
hp journal home
hp labs home
about hp labs
research
news and events
careers @ labs
technical reports
worldwide sites
end of lorem ipsum dolor navigation menu link group
introduction - Volume 8 Number 4

CURRENT ISSUE - Volume 8 Number 4

Jane C. Blake,
Managing Editor

Just 40 years ago, a machine called the TX-0—a successor to Whirlwind—was built at MIT’s Lincoln Laboratory to find out, among other things, if a core memory as large as 64 Kwords could be built. Over the years memory sizes have grown so large that, in the ‘90s, the industry has felt the need to characterize memory in big machines as very large. At five orders of magnitude greater in size than the TX-0 memory, the AlphaServer 4100 8-gigabyte memory is indeed very large, even by today’s standards. Whole databases can be designed to reside in memory. Very large memory technology, or VLM, is a key to the system and application performance discussed in this issue of the Journal, which features the AlphaServer 4100 system, database enhancements from Oracle Corporation and from Sybase, Inc., and extensions to the Alpha architecture.

The AlphaServer 4100 is a midrange, symmetric multiprocessing system designed for industry-leading performance at a low cost. The system accommodates up to four 64-bit Alpha 21164 microprocessors operating at 400 megahertz, four 64-bit PCI bus bridges, and 8 gigabytes of main memory. Opening the section about the 4100 system, Zarka Cvetanovic and Darrel Donaldson describe the project team’s performance characterization of different AlphaServer 4100 models under technical and commercial workloads. Both the process and the findings are of interest. As one example set of data demonstrates, the model 5/300 is not only faster than its DIGITAL predecessors but 30 to 60 percent faster than a comparative industry platform when running memory-intensive workloads from the SPECfp95 benchmark.

The four papers that follow examine areas of the system that challenged designers to keep costs low and at the same time deliver high performance. The AlphaServer 4100 cached processor module design is presented by Mo Steinman, George Harris, Andrej Kocev, Ginny Lamere, and Roger Pannell. Built around the Alpha 21164 64-bit RISC microprocessor, the module is the first from DIGITAL to employ a high-performance, cost-effective synchronous cache, the 4-megabyte B-cache, rather than a traditional asynchronous cache. Next, Roger Dame reviews the clock distribution system, the use of off-the-shelf phase-locked loop circuits as the basic building block to keep costs low, and the signal integrity techniques developed to optimize performance of the clock distribution system for a worst-case clock skew of 2.2 nanoseconds, a goal which the team far exceeded. A unique memory architecture for the model 5/300E is the subject of Glenn Herdeg’s paper. This memory design incorporates a processor module that has no external cache and instead takes advantage of the multiple-issue feature of the Alpha 21164 microprocessor. Closing the section on the 4100 design is the I/O subsystem’s contribution to the system goals of low memory latency and high memory and I/O bandwidth. Sam Duncan, Craig Keefer, and Tom McLaughlin present several innovative techniques developed for the system bus-to-PCI bus bridge design. These techniques include partial cache line writes, peer-to-peer transactions across multiple PCI bridges, and support for large bursts of data.

All efforts to make the hardware run faster are for the benefit of the applications that run on those systems. A paper from Oracle Corporation and another from Sybase, Inc., examine ways in which their respective database systems take advantage of VLM. Vipin Gokhale describes the 64 Bit Option implementation for the Oracle7 relational database system. A primary project goal was to demonstrate a clear performance benefit for two classes of database applications: decision support systems and online transaction processing. The author summarizes data that show a clear benefit for a database with the 64 Bit Option enabled running on the AlphaServer 8400 with 8 gigabytes of memory; in some cases, the performance increase was more than 200 times that of the standard configuration. Sybase engineers T.K. Rengarajan, Max Berenson, Ganesan Gopal, Bruce McCready, Sapan Panigrahi, Srikant Subramaniam, and Marc Sugiyama examine the technology of the System 11 SQL Server that was specifically designed for VLM systems. In addition to achieving record results with the SQL Server design running on the AlphaServer 8400, the engineers have laid the groundwork for future main memory database systems.

Recently, byte and word instructions were added to DIGITAL’s 64-bit Alpha architecture and implemented in the second generation of the Alpha 21164 microprocessor. Dave Hunter and Eric Betts describe the process of analyzing how these additions affect the performance of a commercial database. For testing, the team used prototype hardware, rebuilt Microsoft Corporation’s SQL Server to use the new instructions, and ran the TPC-B benchmark.

The editors thank Darrel Donaldson of the AlphaServer 4100 team and Kuk Chung of DIGITAL’s Database Application Partners group for their efforts to acquire the papers presented in this issue. Our upcoming issue will be devoted to CMOS-6 process and technologies developed by the DIGITAL Semiconductor Division.


Skip page footer
printable version
privacy statement using this site means you accept its terms © 1994-2002 hewlett-packard company
End of page footer
1