February 5, 1993 RM IBM
Beyond workstations
IBM expands its RS6000 workstation line to include multiprocessors at the
high end and a $4,000 diskless workstation at the low end
IBM pushes its RS6000 architecture into areas it previously reserved for
mainframes with vector hardware and supercomputers. A packet-switching
network allows scaling to thousands of processors with constant message
latency on each link.
One upbeat division in a company mired in gloom, the Austin, TX,
workstation group of IBM appears to be on track with additions to the
RS6000 workstation line at the top and bottom, as well as a line of
packet-switching multiprocessors using RS6000 chips that appears scalable
to really large sizes. IBM's confidence in the multiprocessor route to
supercomputing may have much to do with its withdrawal of support for
Steve Chen's supercomputer project. IBM had funded Chen's work after Chen
left Cray. Perhaps most interesting in the IBM rollout is the scalable
Powerparallel System 9076 SP1 (see photo). This number cruncher uses from
8 to 64 RS6000 processors with parallelizing compilers to do such tasks as
computational chemistry and engineering analysis.the limit of the first
box is 64 processors, but a system with 128 processors is already
contracted for Argonne National Laboratory. A synchronous
packet-switching network, with point-to point copper between all
processors, makes the large number of processors able to work together. Of
course, packet switching by itself can't parallelize software to make use of
the hardware, but it does make huge bandwidth available. The network
connecting all the processors uses many identical packet-switching chips
called Vulcan inside IBM. Each switch chip has eight input and eight
output ports, all 9 bits wide (8 data and 1 valid bit). A 1-bit return
token signals receipt of the message. An internal dual-port RAM, 128 x 64,
buffers messages in a queue when more than one is trying to go out the
same port. The network provides a fixed clock frequency to all nodes, one
master and eight offset. In a large system, clock skew cannot be
eliminated, but part of the initialization routine measures the delay in
each chip-chip link. Each receiver latches data at its own edge, but
transmitted signals are clocked out from one of the offset clocks so as to
be in phase at the receiver. A delay of 1 to 15 clocks is acceptable, and
the delay remains fixed on any one link. The switch chip peels off the
first byte as an address and sends the rest of the message, of arbitrary
length, on its way. Any number of bytes could be addresses to successive
switches, so the scheme really will scale to large numbers of processors.
The Powerparallel system runs AIX, and can run multiple jobs on multiple
processors, giving wider use than a machine that can only work with full
parallelism on one program. An RS6000 workstation controls the big box.
Most code for RS6000 workstations can run on the SP1. Parallel code for
the SP1 will be able to run on networks of RS6000 systems.
Workstation upgrades A single-chip implementation of the RS6000 enables
IBM to offer a low-end workstation, the Powerstation M20, starting at $3,
995. The M20 encloses everything in the monitor or the keyboard. It comes
with 16 Mbytes RAM (expandable to 64 Mbytes), a 17-in.color display, a
SCSI interface controller, an Ethernet adapter, one open Micro Channel
slot, AIX, and an AIXwindow license. The M20 runs SPECint92 at 16.3 and
SPECfp92 at 26.7. At the other end of the desktop range, the Powerstation
375, clocked at 62 MHz, rates 59.8 SPECint 92 and 118.2 SPECfp92. With a
6232 CPU (RIOS-1) with an integrated floating-point unit, the Powerstation
375 starts with 32 Mbytes RAM, a 400-Mbyte disk, 19-in. 1,280 x 1,024
color display, and an 80-Mbyte/s implementation of the Micro Channel bus,
all for $25,225. Servers with larger disk and I/O capacities range to over
$100,000. The top of the single-processor line, the deskside
Powerstation/Powerserver 580, with a 3264 CPU, now has a standard CD-ROM
drive and optional tape drive. Its ratings, at 62 MHz like the 375, are
SPECint92 61.7 and SPECfp92 133.2. The extremes of individual
floating-point benchmarks for this machine are 316.5 for 052.alvinn and
57.4 for 077.mdljsp2. The only caches in the machine for these tests were
the on-chip primary instruction (32-Kbyte) and data (64-Kbyte) caches.
The 64-bit RIOS-2 silicon should appear before the end of the year, as
should the Power PC 601. Although IBM and Apple were the prime intended
users of the 601, it appears that the first actual machine using it will
come from SGS-Thomson, in France. –Rodney Myrvaagnes
CAPTION FOR PICTURE OF BIG BLACK BOX:
The SP1 parallel-processing system runs AIX on up to 64 RS6000
processors. The system can run parallelized code for number crunching. It
also runs many separate single processor programs.
Advertisement
