Motivation

Different message passing layers that use the shared network (shared by different applications) were evaluated. TCP sockets performed better than MPICH for the performance indices namely send time and round trip time. This gives a strong indication, that an efficient message passing library can be written using TCP sockets to get a better performance than MPICH. Performance of high-performance interconnection networks together with fast messaging libraries were studied in the evaluation. Fast messaging libraries provide an efficient communication subsystem. This was evident from the results. It imposes a restriction, that communication subsystem performance cannot be increased without the special hardware. The other disadvantage include exclusive access to the network hardware. All these factors lead to the design of TCPMPL.

Some of the features of MPICH that are detrimental to the performance of WARPED are given below:

• Heterogenous Workstations: MPICH is designed for heterogeneous workstations. One part of the application can be run on a SUN SPARCstation, and the other on a Intel Pentium Pro. They can communicate without loss or corruption of information, even though one workstation is in big endian format while the other is in little endian format. This feature was added so that one could maximize the usage of resources when you have a network of heterogeneous workstations. In reality, parallel applications are mostly run on homogeneous machines due to difference in speed and various other factors. To achieve heterogeneity, MPICH converts all the data that is communicated between any two process to a standard external data representation (XDR) [3] format. At the destination, the data is converted back to the local machine format. This can significantly increase the overhead of communication even if the application is executed on homogeneous workstations.
• Functionalities: MPICH is an implementation of MPI standard. The MPI standard offers a myriad of functionalities to encompass all the functionalities needed by different kinds of parallel applications. Some of the functionalities are broadcast, gather, scatter, aggregate and many others. WARPED requires only a few functionalities such as send, receive and probe.
• Network Hardware Independence: MPICH has been designed with ``portability'' as one of the main features. Since MPICH is available for many platforms, parallel application using MPICH can be run on most platforms. To achieve portability, the MPICH has an abstract device interface (ADI) [2] which abstracts the underlying physical communication architecture. There is another level of abstraction called MPID (MPI Device) which make calls to ADI of the specific underlying communication architecture. Finally there is the MPI interface whose calls are mapped onto the MPID calls. This adds a significant overhead if the software does not need to be ported to multiple hardware platforms.

Therefore, limiting a message passing library to provide only the functionality required by WARPED has the potential of increasing the communication subsystem's performance. This lead to the design of a custom message passing library for WARPED to increase the communication subsystem's performance.