Figure 1.

How the Allgather method communicates a spike event by a cell on rank 2 to target cells on rank 4. The illustration uses a maximum integration interval of 1 ms, a buffer size that can hold information about two spikes, and assumes no overflow on any rank during the computation interval between 1 and 2 ms. During the computation interval, a cell on rank 2 fires and its firing time and identifying number are placed in the send buffer and the buffer spike count is incremented. At the end of the computation interval, 2 ms in the illustration, MPI_Allgather sends the contents of its buffer to all other machines (only the receive buffer for rank 4 is shown). Note that the total receive buffer is the size of a send buffer times the number of processes ordered by rank. When MPI_Allgather returns (t is still 2 ms), each process checks all the spikes in its receive buffer by doing a hash table lookup using the cell identifier as key. If a source object for one or more Network Connection objects is returned, the source object with its minimum target destination time is placed on the queue. The computation interval from 2 to 3 then begins.