OS-bypass work is designed to reduce the message latency by allowing the
message-passing layer to talk directly to the network driver software.
It is also designed to allow 0-copy or 1-copy communications between
systems using RDMA techniques.
Since these systems bypass the OS, they must reproduce some of the
functionality that the OS provides. Some method to guarantee
reliable delivery of all the data packets must be found that will still
maintain a very streamlined system.
M-VIA is a modular implementation of the Virtual Interface Architecture
specification that has been developed at Lawrence Berkeley National
Many Fast Ethernet and Gigabit Ethernet cards have been supported.
There is also an
implementation of MPICH called MVICH that is an MPI implementation
running on top VIA hardware and M-VIA software.
An MP_Lite VIA module has been developed to evaluate the performance
of these OS bypass techniques, and to experiment with channel bonding
multiple M-VIA interfaces. Four Fast Ethernet cards were channel
bonded to produce a 3.5 times speedup, while Linux kernel channel
bonding only provides a doubling in performance with 2 cards.
The M-VIA project is no longer funded. The MP_Lite VIA module
has not been used in some time, and may not work anymore.