|
Don Harbert,
Vice President,
UNIX Business
Digital not only invented clusters but continues to set the
standard by which all other cluster systems are measured. The
VAXcluster success and that of Digital's latest UNIX cluster
systems derive from superb engineering that builds on the system
definition put forth in the early 1980s by the VAX engineering
team: an available, extensible, high-performance multicomputer
system built from standard processors and a general-purpose
operating system, with characteristics of both loosely and
tightly coupled systems.[*]
We in the UNIX community are proud of our VAXcluster heritage and
have engineered our products to provide the same kinds of
benefits to customers that VAXcluster systems provide.[+] In the
opening paper for this issue of the Journal, members of the
Digital UNIX engineering team describe the multicomputer system
for the Digital UNIX environment, called TruCluster, which, like
the VAXcluster system, is designed for high availability,
scalability, and performance.
The technology, of course, is different, and the environment is
open. The fundamental concepts are nevertheless the same. The
TruCluster system is a loosely coupled, general-purpose system
connected by a high-performance interconnect. It maintains a
single security domain and is managed as a single system. Cluster
services remain available even when other members are
unavailable. Like VAXcluster systems, TruCluster systems
implement a distributed lock manager, which provides
synchronization for a highly parallelized distributed database
system. The technology for the lock manager, however, is newly
implemented for the UNIX environment. Also completely new is the
interconnect technology for TruCluster systems. MEMORY CHANNEL is
a reliable, high-speed interconnect based on a design by Digital
partner Encore Computer Corporation. MEMORY CHANNEL addresses the
unique needs of clusters by implementing clusterwide virtual
shared memory; the interconnect reduces overhead and latency by
two to three orders of magnitude.[**] Because MEMORY CHANNEL uses
the industry-standard PCI, designers can implement the network at
very low cost. We believe this interconnect technology puts
Digital years ahead of the competition.
The TruCluster system is the latest example of Digital's intent
to remain a technology leader in the UNIX market. We began by
developing the first high-performance, 64-bit general-purpose
operating system, DEC OSF/1, shipping in March 1993. The first
Digital UNIX cluster release, DECsafe Available Server
Environment, followed soon thereafter in April 1994. The
announcement in April 1996 of TruCluster systems with MEMORY
CHANNEL again places Digital far ahead of the competition
technologically. The performance of these available cluster
systems now approaches that of very expensive supercomputers.
System performance has been measured at the record-breaking rate
of 30,390 tpmC on four AlphaServer 8400 systems running Digital
UNIX and the Oracle Universal Server with Oracle Parallel Server.
The previous performance record, 20,918 tpmC, was held by the
proprietary Tandem Himalaya K10000-112; Digital's open system
cluster performance record is 1.5 times the Tandem performance
record at one-third the system cost.
For Digital, clusters of high-performance 64-bit systems are to a
great extent at the heart of its commercial and technical server
strategy. Digital UNIX has been defined and engineered for the
server business, specifically, for the high-performance
commercial and large-problem/scientific environment. To be
successful in the open system market, however, a company must
reach outside itself to jointly engineer products with leading
software suppliers that have the software customers need to be
competitive. Therefore, the first TruCluster implementation is
designed with Digital's partners -- major software companies --
to meet the requirements for high-performance and functionality
in the commercial database server market.
The competitive challenge now is to maintain Digital's
significant lead in providing outstanding cluster performance,
availability, and affordability. From a technological
perspective, the immediate and achievable goal is to increase the
number of cluster nodes from 4 to 10 or 20 nodes. Within this
range, Digital maintains a simple cluster system model that
offers the performance advantages of clustering and avoids the
disadvantages, such as the management problems and qualification
headaches, of more complex topologies. Further, The Digital UNIX
organization will focus on a new cluster file system,
configuration flexibility, management tools, and a cluster alias
that allows a single-system view for clients and peers. The
overall goal of this work is to evolve toward a more general
computing environment.
The kinds of tools that both simplify and enhance performance are
exemplified by the program analysis and optimization tools
presented in this issue. Built on Digital UNIX version 4.0 and
announced in April, these tools help software developers extract
maximum performance from the system. The story of the tools
development is an excellent example of the direct application of
research to products. The power of the OM object modification
tool and the analysis tool with object modification (Atom) was
recognized by developers even as research progressed; in fact,
semiconductor designers developed Atom tools to evaluate new
Alpha chip implementations. The result of this close cooperation
between research and development is advanced programming tools
for customers.
These efforts in the UNIX organization are manifestations of
Digital's commitment to open systems. Other areas of engineering
where this commitment is apparent are also represented in this
issue. For example, eXcursion software is key to integration
between Microsoft's Windows family of products and Digital's UNIX
and OpenVMS products. This wholly revised version both adds new
functionality and conserves system resources. Another major area
of strength for Digital is its networks products. Networks
engineers describe two examples of network services that increase
users' choices and extend system functionality, i.e., the
Integrated Directory Services (IDS) and the Common Directory
Interface.
Digital's strategy is to continue to engineer products that
provide outstanding performance and price/performance in open
environments. In all areas of engineering -- systems, services,
networking -- our goal is to set the standard by which all others
are measured.
* Nancy P. Kronenberg, Henry M. Levy, and William D. Strecker,
"VAXclusters: A Closely-Coupled Distributed System," ACM
Transactions on Computer Systems, vol. 4, no. 2 (May 1986):
130-146.
+ Digital has renamed VAXcluster systems to OpenVMS cluster
systems.
** Richard B. Gillett, "Memory Channel Network for PCI," IEEE
Micro (February 1996) 12-18.
|
|