Comment by dehrmann
The standard rule for CPU-bound RPC server utilization is 80%. Any less and you could use fewer machines; any more and latency starts to take a hit. This is when you're optimizing for latency. Throughput is different.
The standard rule for CPU-bound RPC server utilization is 80%. Any less and you could use fewer machines; any more and latency starts to take a hit. This is when you're optimizing for latency. Throughput is different.
Doesn't this depend on the number of servers, crash rates and recovery times? I wouldn't feel confident running 3 servers running at 80% capacity in ultra low latency scenarios. A single crash would overwhelm the other 2 servers in no time.