Performance

  

Figure: p4 Network Performance

6#6

Efficiency is a major goal of p4, and has been the focus of considerable effort. Although there is still room for further optimization, p4 does not currently add a significant amount of overhead to the underlying transport layer. Although applications are more important than benchmarks, we have occasionally tested p4 in ``benchmark'' mode to make sure that it is reasonably efficient. In Figure we compare the round-trip transmission rates between two processes for a number of message sizes. The curve marked ``ipc/p4'' represents p4 on an ethernet connecting two Sparc 10 workstations. The curve marked ``fddi/p4'' shows that p4 is fully able to take advantage of a faster transmission layer and that an FDDI-connected workstation network can compete with a specialized message-passing machine. The curve marked ``i860'' shows message-passing performance of the same program converted to run with Intel primitives. (The barely visible ``zigzag'' at 128 bytes shows that the test program is sensitive enough to pick up the protocol change that occurs on the iPSC/860 at that message length.)