The CM-2 data transposition problem

Abstract

The CM-2's natural data layout is not conducive to exchanging data with other machines. Before CM-2 data is sent to a remote machine, a bitwise transpose must be performed on the data. Each bit in an n bit value must be transmitted to a different processor, requiring n send operations through the CM-2's global router network. The time required to transpose the data limits the effective throughput of the I/O channel to a small fraction of its peak theoretical bandwidth. For example, when sending data to a remote supercomputer using a 100 MB/s HIPPI channel, an effective throughput of only 4.9 MB/s can be achieved. The authors describe the CM-2 transpose problem and study ways to improve the performance of transposed data transmissions.<>