MPI applications need an initialization procedure that create processes and prepares necessary information for passing messages each other. Such procedure depends on the functions given by middlewares. For example, MPICH-G2 use DUROC, which is a component of Globus, to initialize MPI processes. Sometimes it can be very efficient, but it makes the MPI library to be tied up by the middleware. Thus, the change of Globus results in inevitable change of MPI library. This is the reason why the MPICH-G2 cannot work with Globus 3.0 or later.
Therefore, we modify the initialization procedure of MPICH-G2 which depends on DUROC component to file-based initialization. It allows users to launch MPI applications without DUROC. To use file-based initialization, users have to make the file (we call it configuration file). JobSubmission Service in GRASP helps you to run your MPI applications in convenient way. It can create configuration files and stage the files to execution node automatically. But, you also can launch the applications manually without any launching program if you know the method to make the configuration file. The contribution of the file-based initialization is that it is independent of middlewares. Even though the Grid middleware is changed more and more, we can launch the MPI application in a same way when we use the file-based initialization.

It is a well-known problem that MPICH-G2 does not support private IP clusters. In MPICH-G2, processes communicate with each other based on IP addresses. Thus, it cannot support private IP clusters by the nature.
To support private IP clusters, MPICH-GX uses a communication relay scheme combining the NAT service with a user level proxy named ‘nproxy.’ In this approach, only incoming messages are handled by a user-level proxy to relay them into proper nodes inside the cluster, while the outgoing messages are handled by the NAT service at the front-end node of the cluster. Figure 2 show the conceptual diagram of our relay scheme.

For the collective operations, MPICH-G2 uses the flat-tree algorithm which the root communicates with other processes directly (See details in http://www.globus.org/mpi) But, when if the high latency occurs between the rood and other processes, its performance can be degraded significantly. Thus, MPICH-GX uses the HLOT (Hierarchical Latency Optimal Tree) algorithm for collective operations to avoid such high latency links.