Today I did a comparison between the textual and the binary protocol.
System calls:
I used a single thread to insert 10 000 object into the cache and then
perform 10 000 get operations. I used truss -c to count the number of
system calls on the server, and the textual and the binary protocol
used the exact same number of read / sendmsg calls.
Performance:
I used my test-program to simulate 8 active clients in a pool of 5000
clients (using 8 threads and that picks a random idle client and
performs a get-operation). I did 1 000 000 get-operations for 1k
objects, and the performance and CPU usage was pretty much the same
(The binary protocol a little bit (2-3%) better/faster).
Trond
