HERSON@SRI-ROBOTX.ARPA (Jim Herson) (12/10/87)
We are using DECNet for simple task to task communications and
are very disappointed in throughput. We are wondering if we
have implemented something incorrectly. Our benchmarks are as
follows:
time per | machine running | machine running
packet(ms) | process A | process B
---------------------------------------------
4.0 | 8600 | 8600
---------------------------------------------
28.3 | 8600 | 730
---------------------------------------------
34.7 | 730 | 730
These times are based on an average of 1000 packets, each 500 bytes long.
3/4 of the total time is spent in the network and 1/4 in our code. All
tests were run on unloaded machines (each with plenty of memory). The
basic loop is:
1. issue a read request
2. hibernate until the read completes and the AST is invoked.
3. do some minor bookkeeping and turn the packet around.
Any pointers or information on similar experiences would be appreciated.
Thank you
Jim Herson
(HERSON@ROBOTX.SRI.COM)
-------