The proposed network provides one-to-many communications as well as one-to-one communications wave-pipelining (WP) method. Only when a bus is required, the network virtually establishes a bus. We name it "virtual bus". If urgent message or multi/broadcast message is occurred, it can intervene on-going one-to-one communication by freezing the proceeding of on-going message. In this scheme, network interface(NI) shares data lines with control lines. Thus the proposed scheme is resource-efficient. The virtual bus system is superior to the real bus system in two respects: 1) better bandwidth utilization, and 2)the low latency communication with higher bandwidth. We adopt wave-pipelining technique into the NI design. WP can be four times faster than conventional pipelining[1]. For integrating it to NI, we proposed automatic skew sampling circuit that eases the implementation of a wave-pipelined NI[2]. The skew sampling circuit detects the delay differences between parallel signal lines, samples each signal, and merges the signals to have the same phase. And we ported MPI1 ＆ MPI2 for virtual bus network (MPI/VBus). By using user-level communication, we could reduce the software overhead. MPI/VBus and device driver for virtual bus share the send/put/get queues in kernel area so that we could reduce system call of MPI/VBus, ie, MPI/VBus only to enqueue messages when the queues are not empty. MPI/VBus uses of internal broadcast and multicast operations of virtual bus network so that there are no multiple unicasts for broad/multicast request.
