|
|
BenchADM (80-cubed)
The BenchADM benchmark
application is the computational kernal representative of
many applications in numerical relativity. It is
characterised by being compute intensive, with around 600
Floating point operations per grid point and with six grid
functions synchronised each timestep. The benchmark is run
here with a local workload of 80-cubed on each processor).
| Machine ID |
Architecture |
Fortran Compiler |
Single Processor
(secs) [1] |
MFlops [2][3]
(%Peak) |
Scaling
(16 procs) [4] |
| IA64 Linux |
|
| Titan |
IA64
VMI/GigE |
Intel 6.0 (-O3)
(20020320) |
123.6 |
228.3 (7.1%) |
89%
data |
| Titan |
IA64
VMI / Myrinet |
Intel 6.0 (-O3)
(20020320) |
123.9 |
227.7 (7.1%) |
94%
data |
| IBM |
|
| Psi |
Power4 |
xlc (-O2) |
Interact.: 61
RZG: 55 |
550 (11%) |
52%
data |
| NEC |
|
| hwwsx5 |
NEC SX-5 |
native |
30.9 |
457 (23%) |
62% (on 4 procs)
data |
Notes
- Measured using Cactus timer (Total time) with gettimeofday clock.
- MFlops: This is calculated using the number of floating
point operations calculated on the Origin 2000 using perfex
including standard optimisation but
switching off multadds
(compiling with -TARG:madd=OFF). The number of floating
point operations for the different cases are listed on
the pages describing the individual benchmarks (to come).
- MFlops here means 1,000,000 Flops.
- Scaling is defined by
user time on one processor/user time on 16 processors
|
|
|
![[MAINIMAGE]](../../images/development.jpg)
