> Right. It does sound like we'll have to conduct some experiments. > Would've been nice to get some other inputs though. > > We might make 2000-3000 mgets/sec. 100-500 keys each. >
It's straight up math. if (keylength * keycount > tcp_buffer) -> latency benefit from having shorter mgets (or just splitting them up and sending them to the same server in parallel). Otherwise it's a syscall count for the # of servers you hit. It's actually really hard to determine at a glance which will work better. Clients implement this part differently, so it may just be up to which they implement better.
