Machine type | RISC-based distributed-memory multi-processor |
---|---|
Models | AlphaServer 8200 5/440, 8400 5/440, Cluster |
Operating system | Digital Unix (DEC's flavour of Unix) |
Connection structure | Crossbar (see remarks) |
Compilers | Fortran 77, HPF, C, C++ |
Vendors information Web page | http://www.digital.com:80/info/hpc |
System parameters:
Model | 8200 | 8400 | Cluster |
---|---|---|---|
Clock cycle | 2.3 ns | 2.3 ns | 2.3 ns |
Theor. peak performance | |||
Per proc. (64-bit) | 875 Mflop/s | 875 Mflop/s | 875 Mflop/s |
Maximal (64-bit) | 5.25 Gflop/s | 10.5 Gflop/s | 84 Gflop/s |
Main memory | <=6 GB | <=14 GB | <=36 GB |
Memory bandwidth | |||
Processor/memory | 1.6 GB/s | 1.6 GB/s | 1.6 GB/s |
Between cluster nodes | --- | --- | 100 MB/s |
No. of processors | 6 | 12 | 96 |
Remarks:
The AlphaServers are symmetric multi-processing systems which are based on the Alpha 21164A processor. The 8200 model is a somewhat smaller copy of the 8400 model: in the 8200 a maximum of 6 CPUs can be accomodated while this number is 12 for the 8400 model. Also, there is room for at most 6 GB of memory in the 8200 while the 8400 can house 14 GB. However, the amount of CPUs and memory is not independent. For instance, the 8400 has 9 system slots. One of these is reserved for I/O and one will have to contain at least one CPU module which can contain 1 or 2 CPUs. From the remaining slots 6 can be used either for memory or for a CPU module. So, one has to choose for either higher computational power or for more memory. This can potentially be a problem for large applications that require both.
AlphaServers can be clustered using PCI bus Memory Channel link cables that are connected to a hub. The systems need not be of the same model. The bandwidth of this interconnect is slightly over 100 MB/s. Up to eight systems can be coupled in this way. To support this kind of cluster computing, HPF and optimised versions of PVM and MPI are available.
Measured Performances: As yet, only single system results for the AlphaServer 8400 are available. No cluster results are known. In [4] a speed of 6.7 Gflop/s for an 12-processor system are reported for the solution of a dense linear system of order 9548.