PVFSv2 is a redesign of the first PVFS that improves the modularity of the implementation. The BMI layer of PVFSv2 was designed to give the needed network abstraction and messaging layer. This work presents a new BMI module that supports the VIA. The shows that the BMI requirements and semantics for messaging were accomplished along with several other design goals. The implementation successfully operates on both Myrinet and InfiniBand. It implements a credit based flow control mechanism to fulfill one of the BMI requirements and allows for benchmarking and examination of its performance. The baseline bandwidth and latency of the implementation were compared to the BMI modules and were shown to achieve significantly higher performance than the TCP module, but slightly less than the GM module. The CQueue version of the implementation offers the best overall performance when running over the VI-GM implementation. Test results show that the InfiniBand Verbs used to implement VI-IB are not as mature as the GM protocol for Myrinet used in VI-GM. At the time of this study, the VIPL from Intel was meant as an early reference design and may not be fully optimized. BMI is a network abstraction layer well suited for systems-level programmers. Most BMI users will be using the API to build flexibility in their parallel system such as PVFSv2. If each BMI application that uses the implementation knows the most prevalent size of messages that will be sent, it should configure the implementation at runtime to use the eager protocol for this message size. In general, immediate messages offer better performance than rendezvous messages, especially if the VIPL is implemented using another low-level protocol such as GM. If the immediate/rendezvous threshold is going to a large message size, the send the credit limit should be also configured at runtime to a smaller number to decrease memory requirements of implementation. It was also shown that a low send credit limit combined with a large IR-threshold has a negligible impact on performance. The Notify mechanism was only able to out-perform the CQueue version in one benchmark. This was an unexpected result, because having routines handle the incoming message immediately would seem to offer better throughput. However, the context switch time and the overhead of multiple threads decreased throughput.