This system is similar to CMMD on the CM-5, but it has
no synchronous calls (such as CMMD_send_block()).
We have seen that on the CM-5, CMMD_send_noblock() is clearly less efficient
than the other sending functions.
On the Paragon, the performance of csend() is
as high as the performance of isend() on a ``ping-pong" test.
In fact, in this kind of test,
the receive is always posted when the data is to be sent. Thus,
no extra buffering occurs. The Paragon in this respect is more efficient
than the CM-5, which always does an extra buffering (see 
).
In the following experiments, we used csend()-crecv() to asses the native bandwidth and latency.