J Fieb, Here is the article:
November 24, 1997, Issue: 982
Section: Technology
Distributed computing highlights
Supercomputing 97 -- Networks harness
supercomputers
By Chappell Brown
San Jose, Calif. - The quest to build the ultimate supercomputer is showing signs
of becoming enmeshed in the networking revolution. As the technical program at
last week's Supercomputing 97 (SC97) show revealed, clusters of
high-performance workstations linked to central supercomputing resources may
prove to be the ultimate problem-solving tool.
New approaches based on high-performance networking, or simply on the global
reach of the Internet, reveal some of the potential that widely distributed
computing systems offer.
Networking already addresses the need for scientists and engineers to collaborate
on projects. Whether it is a group of electrical engineers within a company
designing a system or a group of physicists dispersed over geographically remote
universities, sharing data and algorithms is critical to large-scale efforts. Networks
allow such groups to tap the latest supercomputers when needed without having
to make a major investment in hardware.
"I think a lot of people are realizing that they can't afford the latest supercomputer
and a more-economical route is to link smaller resources over networks," said
Andrew Chien, a computer scientist at the University of Illinois at
Urbana-Champaign, who presented his work on networks for high-performance
clusters at the conference. "Networking has really taken over in the
supercomputing world and you can see the extent of it at a conference like this,"
he added, pointing out that commercial vendors such as Hewlett-Packard Co. and
Sun Microsystems Inc. are introducing clustered systems and high-speed
interconnect to reach supercomputer levels of performance.
Chien's concept, evolved with colleague Kay Connelly at the University of Illinois,
is to simplify the network for high performance and ease of use. "We would like
to be able to aggregate resources easily while maintaining network quality." The
design involves a simplified switch for routing data along with a higher-level
network-management system that ensures the timely delivery of data between
nodes.
Called FM-QoS the network has a specially designed traffic-management system
that can predetermine the allowable time delays between two nodes. The approach
uses feedback from the destination node to the node of origin to maintain a given
level of performance. Quality of Service (QoS) is the term that is generally used
to express the need for a specified level of data throughput for a network. The
QoS issue arises because different computational tasks have different
requirements for the delivery of data. A high-speed numerical simulation might be
able to tolerate only millisecond delays in data delivery while a group of users
interacting online could absorb much longer delays. The FM-QoS system insures
that each process can operate at its unique speed requirements without having
network traffic interfere and block program execution.
By using network feedback with self-synchronizing communication schedules,
the network-management system achieves synchronous I/O at all network
interfaces. The network can therefore be scheduled for predictable performance
without needing any specially designed QoS hardware. A prototype system has
been demonstrated over Myricom's Myrinet network that showed synchronization
overheads of less than 1 percent of the total communication traffic. The
demonstration network achieved predictable latencies of 23 ms for an eight-node
network, which Chien said was more than four times faster than conventional
best-effort schemes that could achieve a predictable delay of only 104 ms.
The current design is optimized for clusters of workstations and supercomputers
in a building and there is flexibility in the architecture to scale up to large systems
such as a campuswide computer cluster. "It resembles the structure of 'clock
domains' on VLSI chips where subcircuits are carefully timed and then linked by
a clock that can have more skew," Chien said.
At the other extreme is the Internet, which is not under local control, but
nevertheless offers opportunities for solving supercomputer-level problems. Some
breakthroughs in solving intractable problems have resulted from a group of
researchers breaking the problem down into smaller segments, doing computer
runs locally and then combining the results. Systems such as the Parallel Virtual
Machine, a network operating system that emulates a supercomputer in terms of a
set of widely distributed computing resources, are being evolved by engineers at
supercomputing centers.
That technique may get a boost from an Internet-based operating system called
the Scalable Networked Information Processing Environment (Snipe) being
developed at the University of Tennessee. Collaborations over the Web involve a
lot of "hand coding" in the sense that everyone involved needs to take their
segments of the problem and code them locally on whatever computing resources
they have. The system would automate that process by identifying and tracking
algorithm segments and data that are distributed over the Web. "We started with a
Web model in which all the segments of a process are a resource tagged with a
URL [uniform resource locator]," said Keith Moore, who presented the latest
results from the project at the conference. Not only is the Internet unique in the
scope of its interconnectivity but it can potentially involve millions of users. The
Snipe system therefore duplicates data and algorithms and dynamically allocates
them to computing resources on the Web.
A local cluster of computing resources-workstations, supercomputers and
high-performance database servers-would be assigned a URL. The Snipe system
then sets up virtual interconnects between these clusters and implements them
over the Web, managing data transfers and resource duplication to emulate the
virtual network of links.
Evaluating traffic
A variety of other network design and evaluation tools surfaced at the conference.
Systems for evaluating traffic and performance at the network, I/O and inside a
given computer system were described in several technical sessions. Clusters of
symmetric multiprocessors and virtual memory-management systems over
networks seem to be emerging as a force in high-end computing, answering the
need for tightly managed local computing resources that support work groups.
One system that may help diverse technical computer users make sense of their
execution environment is a performance evaluation tool from the University of
Wisconsin-Madison, presented by Karen Karavanic, one of the researchers
developing it. Called the Experiment Management System, the program is
organized around the metaphor of a lab notebook. The system records results
from a set of program executions collected over the life of an application. The
data is stored as "program events," which include the components of the code
executed, the execution environment and the performance data collected.
The combinations of code and execution environment are organized as a
multidimensional program space. Visualization methods allow the user to view the
collected data using different criteria. A graphical representation of the program
space is the basis for the interface to the Experiment Management System.
"Essentially you want to answer the question 'how and how much did the
performance change?' " Karavanic said. Currently, complex and detailed
benchmarking procedures are used to answer that question, but those systems are
relatively fixed. The lab-notebook system is designed to make the whole
evaluation process more flexible by including it as part of an ongoing
computational project.
"For example, we tried the system out with a group of chemists that were doing
computer-based chemical simulations. The evaluation process was compressed
from about a year down to a few weeks," Karavanic said. The lab-notebook
metaphor helps scientists, who may not be computer scientists, to perform
effective performance evaluations as part of their scientific experiments. "This
system is not specifically tied to computer-system evaluation," said Barton Miller,
another researcher on the project who evolved the basic concept. "We intend to
develop it as a general lab-notebook system that could be used in any lab."
The computerized notebook includes the chronological page-by-page structure
required by standard lab practice. But the same information, once recorded, can
be reorganized by the system to display the results in a variety of modes to help
high-performance-computer users to grasp the effectiveness of their particular
mix of applications and systems. |
