Story by Jenn Wagaman
Launched in 1993 with a 4-CPU Cray YMP M98 supercomputer, a 1.1
terabyte StorageTek™ Powderhorn tape silo, and a dedicated
staff of 11, the Arctic Region Supercomputing Center (ARSC) embarked
on a bold plan to establish a first-class supercomputing facility
in Alaska. Located at the University of Alaska Fairbanks, the center’s
focus is to support high performance computational research in
science and engineering with emphasis on high latitudes and arctic
regions encompassed many scientific disciplines.
In the past decade, the number of ARSC staff increased to over
fifty, in addition to a half dozen ARSC joint faculty members from
the UAF departments of computer science, engineering and art; several
student researchers and summer interns; vendor support analysts;
and a number of affiliated faculty and visiting professionals.
As the first year of the center’s second decade unfolds,
users have access to the computational power of three Cray and
two IBM supercomputers, two StorageTek™ robotic tape silos
capable of storing a petabyte of data and a variety of visualization
resources. In addition, a new Cray X1 arrived in the summer of
2003. A new IBM Power4 cluster system arrived in October 2003 and
is currently undergoing installation. Two Sun Fire™ 6800s,
added late in 2002 to support mass storage, are planned for full
production by fall 2003.
Extending the opportunities for researchers to visualize and explore
their data in an immersive three-dimensional environment, ARSC
added a Mechdyne Flying Flex™ three-wall and single-floor
display system to further enhance scientific discovery processes.
Currently, researchers use ARSC resources to solve problems in
global climate change, space physics, ocean circulation, galactic
formation, computational fluid dynamics, bioinformatics, arctic
engineering and art, in support of academic and High Performance
Computing Modernization Program (HPCMP) allocated projects.
Computational Upgrades
Klondike
Cray X1™ In May 2003, ARSC installed the first phase of the Cray X1™ 128-processor
parallel vector system, named Klondike, which is the successor
to Cray’s vector and MPP systems.
With Cray’s new streaming innovation, 128 vector processors
give a parallel programmer 512 very powerful processors on which
to run codes. This is accomplished using multi-streaming processors
(MSP), comprised of four single-streaming processors (SSP). Each
MSP is capable of 12.8 GFLOPs of theoretical peak performance for
a total of five teraflops for the system.
“
This machine is going to provide a new opportunity for blending
the two paradigms,” says Ed Kornkven, ARSC vector specialist. “You
really have to look below the surface to see all that’s going
on in this machine.” The multi-streaming is useful in situations
when the compiler can see the opportunity for further dividing
of vector work and allocating it to the SSPs, thereby compounding
vector capabilities with parallelism.
Programs that run well on the Cray SV1ex™ and can take advantage
of vectorization, multi-streaming and parallelism should do well
on the X1™ architecture. Through its combination of vector
processors with distributed shared memory (DSM), Klondike is able
to support OpenMP programming within an MSP and MPI programming
using C, C++ or Fortran across nodes. Co-Array Fortran and UPC
can also be used for parallel programming.
ARSC’s X1™ provides a theoretical peak performance
of 1.6 teraflops and will eventually replace the center’s
current Cray T3E™, Yukon, and Cray SV1ex™, Chilkoot.
Pioneer users were added to the system in October while final testing
took place. General availability of the system began in early November
2003.
Iceberg
IBM p690+/p655+/p655 In the Fall of 2003, ARSC added an integrated-architecture IBM
supercomputer system named Iceberg, composed of two IBM p690+ systems,
each with 32 processors and 256 GB of memory. Iceberg also includes
92 IBM p655+ systems, each with eight processors and 16 GB of memory.
In addition, there are p655 interactive login and I/O nodes, each
with eight processors and 16 GB of memory per node. Running the
AIX operating system, this system integrates with IBM’s next
generation clustering technology. The overall peak theoretical
performance is five teraflops. All of the processors in Iceberg
are IBM’s Power4 architecture, compatible with Iceflyer,
ARSC’s 32-processor IBM Regatta.
The two p690+ servers are configured as 32-way shared memory
systems making them ideal for large-memory OpenMP jobs. The bulk
of the
computing capability is in the p655+ servers. These can be used
as small eight-way shared-memory systems, and can also be used
for large message passing jobs using multiple nodes per job. The
batch system on Iceberg is loadleveller, as on the other ARSC IBM
systems.
A new feature in Iceberg is the density of the p655+ nodes and
the interconnect. This architecture is the first time IBM has put
128 CPUs in one air-cooled frame. Inter-node communication is handled
by the new generation Federation Switch, replacing the old SP and
SP2 switches and offering extremely low latency and very high communication
speeds. Iceberg will be the largest ARSC IBM system.
|
