SC2000 The exhibitionUtrecht 30 November 2000 Although the exhibition floor was not noticeably larger that in the last few years, the amount of exhibits was certainly larger. The reason for the growing activity can be found in the large number of small startup companies that try their hand in cluster computing, grid computing, or what was dubbed Megacomputing in the Panels section. There was a variety of firms that sell preconfigured Beowulf clusters, the vast mojority based on Linux, some based on Window, and one (BTG) that offers a SparC-based cluster with the Solaris OS for Beowulf prices.
Another part of the cluster-related vendors is offering software for
managing clusters and adding other software to enhance the
functionality of the standard software distributions that circulate for
Beowulf clusters (one is even called MPI).
All important vendors now also offer clusters except, for the moment,
HP. Sun goes through BTG to market SparC clusters, Compaq offers Alpha
and Intel-based clusters, Fujitsu-Siemens, IBM, and SGI also sell
Intel-based clusters. In other words, the traditional High Performance
Computer vendors all have recognised that clusters are a lasting
phenomenon and they want their share at least to complement the
integrated HPC systems that were marketing over the years. Most vendors
were showing there well-known products or natural evolutions of these.
To name a few: IBM showed its POWER 3 based SP system. Last year already
a system was available with 4 CPUs in an SMP node. This trend has been
continued: the new high nodes can harbour up to 16 375 MHz POWER 3 CPUs
now and also the new switch between the nodes has become roughly 3 times
faster. Similarly, SGI is now offering the Origin3000 line with faster
400/500 MHz MIPS R14000 processors and a network that has a nominal
speed that is more than 2 times larger than in the earlier Origin2000,
also the system has been more modularised, making it easier to configure
it to the customers needs.
Cray Inc. is the merger of the former Tera
company and the part of SGI/Cray that manufactures vector systems. At
SC2000 they announced the Cray SV-1ex. A mid-life kicker system with a
50% increase in speed with respect to the SV-1 with a peak speed of 1.8
Gflop/s per vector stream. Furthermore, the complete CMOS implementation
of the MTA multi-threaded system is expected to be out in April 2001.
HP presented its SuperDome server with 4 CPU nodes called ``cells''
which in turn are connected by a crossbar to up 16 cells. The SuperDome
is a ccNUMA system, like SGI's Origins: all data in the system is shared
but not available on the same time scale. The CPUs are PA-RISC 8600
processors at 552 MHz with a Theoretical Peak of 2.2 Gflop/s.
Both NEC and Fujitsu were offering systems that were not based on vector
processors. Fujitsu featured its PRIMEPOWER 2000 systems, a natural
evolution of its GP7000 SparC-based general servers. The PRIMEPOWER can
harbour up to 128 SparC64 GP processors at 450 MHz and looks more geared
for transaction processing than being targeted to the HPC community. NEC
showed it 16-way AsuzA system which is based on Intel's Itanium
processor. The system runs under Linux and is built from 4-CPU cells
that communicate via a separate data crossbar and address network. As
the Itanium processors are not yet in regular supply, the systems will
be shipped in the first half of next year at the earliest.
One system presented was completely new: the SRC 6. SRC Inc. claims to
be the direct heir of Seymour Cray's Cray Computer Corporation. The
system they are presently building consists of 256 2-CPU Intel boards
that communicate via separate read and write crossbars with a common
memory while each board is enhanced by a so-called MAP, or
Multi-Adaptive Processor. In fact the MAP consists of an FPGA with
on-board memory and access to the general processors on its board and
the memory. The MAP can be configured for user-specified like data
compression or encryption, signal processing, rendering, etc. These
special purpose enhancements may be much faster than what can be
achieved with the general processors but has to be configured
separately in the MAP, making it a hybrid between a general and a
special-purpose machine.
Aad van der Steen |
 |