MPI for SCI-connected Clusters

Project Description

Since the presentation of the first standard in 1994, the Message Passing Interface (MPI) has become arguably the most important API for parallel high-performance computing. The freely available Open-Source implementation MPICH was very important for MPI's success. MPICH is the most commonly used, freely distributed implementation of the MPI-1 standard (including parts of the MPI-2 standard) which is also used as a basis for commercially distributed MPI Implementations.

MPICH supports different communication mechanisms for exchanging messages between the cooperating processes in a parallel application, e.g. shared memory and TCP/IP. Today, the most popular architecture for parallel systems is having multiple PCs connected via Gigabit Ethernet. The standard MPICH distribution can be used to run MPI applications on such a system.

However, to get better performance, a high-speed local network providing lower latency and higher bandwidth than the standard Ethernet can be used. Such more sophisticated interconnects are not supported by the standard MPICH distribution, but often by some MPICH derivative. SCI-MPICH is an example for this. It consists of an ADI-2-Device (an implementation of the MPICH driver layer) which supports the Scalable Coherent Interconnect (SCI).

I'm afraid that our terminology regarding SCI-MPICH might sometimes be a bit confusing, so here are the main terms with a short explanation:

• MP-MPICH is the name of the development project consisting of SCI-MPICH, NT-MPICH and MetaMPICH as well as the name of the software distribution which we provide as an outcome of this ongoing project.
• SCI-MPICH is the name of the subproject of MP-MPICH in which we develop the SCI support for MP-MPICH. An MP-MPICH distribution configured for usage with SCI might also be called SCI-MPICH.
• ch_smi is the name of the ADI-2-Device that supports the usage of MP-MPICH over SCI. ch_smi is the software that is developed in the SCI-MPICH project.
• The name ch_smi stems froms the underlying SMI library. The SMI library allows for cluster-wide shared memory programming on an SCI cluster. It is based on the SISCI library from Dolphin Interconnect Solutions.

The SCI-MPICH project is finished. The last version of the software is contained in MP-MPICH 1.5.

News Archive