Ames recently purchased an SGI 512 CPU Origin
2000 system. The system has been named Lomax,
after the late celebrated Ames researcher Harvard
Lomax. The Lomax system is the largest single
shared-memory multi-processor system in the world
(see figure 1). It is the result of an Ames-driven
partnership with SGI to push the limits of single-system
shared memory designs. It is believed that
large CPU count single-system designs offer many
potential advantages in those research areas that
require very high levels of parallel computational
performance. This system has demonstrated over
60 billion floating-point operations per second
(60 GFLOP/sec) of sustained performance for the
production computational fluid dynamics (CFD) code
OVERFLOW-MLP (13 times that of a 16 CPU C90
system). This system offers even higher performance
potential for molecular dynamics simulations.
Recently, the Lomax system was used as the
parallelization testbed for the COSMOS ab initio
molecular dynamics model used in NASA's astrobiology
research effort. The COSMOS code is often used
to perform protein-folding simulations. Historically,
many important problems involving 20,000-30,000
atoms have not scaled well on "clustered" parallel
systems. This lack of performance is due to the small
amount of work performed by each CPU relative to
the time spent transferring data between CPUs.
The single-system approach of the SGI Origin
2000 architecture, and the large CPU count Lomax
system in particular, offers an ideal platform for such
computations. The Origin design supports very fast
and low latency memory access times from any
processor to any memory module. This low latency
and high performance are essential for parallel
scaling to the hundreds of CPUs necessary to execute
problems in a timely manner.
The optimization effort is focused on inserting the
highly efficient Ames-developed multi-level parallelism
(MLP) approach into COSMOS. At this point the
two major time-consuming routines have been
converted with highly encouraging results. The first
routine computes its zones between all water molecules
in the system (WATNLS1). The second
(MPFGATHER) gathers the forces for subsequent
molecular movement. The results are summarized in
Table 1.
| Table 1. A comparison of COSMOS and COSMOS-MLP execution times. |
| COSMOS (32 CPUs) | COSMOS-MLP (343 CPUs) |
| Module Summary | Module Summary |
| WATNLS1: 56.66 | WATNLS1: 0.94 ( 60x) |
| MPFGATHER: 42.13 | MPFGATHER: 0.11 (383x) |
| BARRIER: 0.08 | BARRIER: 1.97 |
| Totals: 98.87 | Totals: 2.92 ( 36x) |
As the table shows, the MLP modifications
dramatically improve the code performance on the
two most time-dominating routines. The speedup
arises from the much higher scaling efficiencies
found in the MLP based parallel algorithm, coupled
to a greater reuse of encached data. It is this
expanded cache reuse that fuels the observed
dramatic superlinear speedup over the old code
executing at its parallel limit of 32 CPUs.
Current efforts indicate that COSMOS-MLP
executions on Lomax will be some of the fastest ever
achieved in this field. The results of this research
have far-ranging implications in the commercial
world, for the advanced numerical techniques
developed under this effort are generally applicable
to a number of industry standard models used by the
university and drug research communities in the
United States.
Point of Contact: J. Taft (COSMOS-MLP)/A. Pohorille (COSMOS)
(650) 604-0704/5759
jtaft@nas.nasa.gov
pohorille@raphael.arc.nasa.gov
Back To Top
Previous Paper
Return to Revolutionary Technology
Next Paper 
