Next: UTLB overheads
Up: Impact on Data Transfer
Previous: Impact on Data Transfer
Figure 4 shows the
one way latency while Figure 5 shows the bandwidth of
synchronous remote send operations on the two platforms. The latency
is measured using a pingpong benchmark in which two processes on two
different machine send synchronous messages to each other back and
forth. The bandwidth is measured using a pair of processes on two
different machines where one process continuously does synchronous
sends to the second process. Because the PCI bus bandwidth(133 MB/s)
and myrinet bandwidth(160 MB/s) are very high, the dominant portion of
time of small message communication is the processing overhead on the
host and the network interfaces. The one-way latency for small
messages is 14 s while the peak bandwidth is about 96 MB/s on both
platforms.
The VMMC communication model also provides a remote-read capability to
grab data from the memory of the remote machine without the
involvement of host processor that machine. Figure 6
and 7 show the latency and bandwidth performance
respectively. Both measurements are performed using a pair of
processes on two different machines where one process repeatedly
performs reads from the second process' memory using the remote-read
operation.
As expected, we see that the OS has little impact on the data transfer
in the common case. Later optimization on NT further reduces
the latency by 6 s after our port.
Figure 6: Remote fetch latency
Yuqun (Michael) Chen
Wed Jun 2 19:35:36 EDT 1999