Next:
The Cray MTA.
Up:
Recount of (almost) available ....
Previous:
The Bull NovaScale.
| Machine type |
RISC-based distributed memory multi-processor. |
| Models |
C-DAC PARAM Padma. |
| Operating system |
AIX (IBM's Unix flavour), Linux |
| Connection structure |
Clos network. |
| Compilers |
Fortran 77/90, C, C++ |
| Vendors information Web page |
http://www.cdacindia.com/html/parampma.asp |
| Year of introduction |
2003. |
System parameters:
| Model |
C-DAC PARAM Padma |
| Clock cycle |
1 GHz |
| Theor. peak performance |
| Per Proc. (Gflop/s) |
4 |
| Maximal (Gflop/s) |
1024 |
| Memory |
500 GB |
| No. of processors |
248 |
| Comm. bandwidth |
| Aggregate |
4 GB/s |
| Point-to-point |
312 MB/s |
Remarks:
The PARAM Padma is the newest systems made by the Indian C-DAC. It is built
somewhat asymmetrically from 54 4-processor SMPs and 1 32-processor node. All
nodes employ 1 GHz IBM POWER4 processors. As an interconnection network C-DACs
own PARAMnet-II is used for which a peak bandwidth of 2.5 Gb/s (312 MB/s) is
given with a latency for short messages of ≅ 10 µs. The network is
build from 16-port PARAMnet-II switches and has a Clos64 topology, very similar
to the structure used by Myrinet. No MPI results over
this network are available.
C-DAC has already a long tradition of building parallel machines and it has
always provided its own software to go with them. Therefore, the Padma comes
with Fortran 90, C(++), MPI, and a Parallel Files System.
Measured Performances:
The Padma performs at 532 Gflop/s with the HPC Linpack Benchmark (see [44]) for a linear system of size N =
224,000 on a 62-node machine with a theoretical peak of 992 Gflop/s. That
amounts to an efficiency of 53.6% for this benchmark.
Next:
The Cray MTA.
Up:
Recount of (almost) available ....
Previous:
The Bull NovaScale.
Aad van der Steen
Tue Mar 8 14:09:42 CET 2005
|