4 ms·
> People ... aren’t machines that can run constantly on 100% utilization. You also can't run machines at 100% utilisation & expect quality results. That's when
by dwattttt 2y ago
> People ... aren’t machines that can run constantly on 100% utilization.
You also can't run machines at 100% utilisation & expect quality results. That's when you see tail latencies blow out, hash maps lose their performance, physical machines wear supra-linearly... The list goes on.
- dehrmann 2y agoThe 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.
- pdhborges 2y agoDoesn'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.
- dehrmann 2y agoRight; this is only for large pools of servers.
- namaria 2y agoDifference is machines break and that costs lots of money. People just quit, some businesses consider it a better outcome.