Hacker News Reader

Comment by throwaway4good | Hacker News Reader

I would expect most high-end servers in data centers to sit idle most of the time? Do you know of any data on this?

Most servers are doing things for human beings, and we have irregular schedules. Standard rule of thumb is that you plan for a peak capacity of 10x average. A datacenter that doesn't have significant idle capacity is one that's some kind of weird special purpose thing like a mining facility.

That's true, but I would expect that most idle servers are turned off and they use Wake-on-LAN to become active when there is work to do.

Just a few servers could be kept idle, not off, to enable a sub-second start-up time for some new work.

Servers are very expensive. It is not cost efficient to buy a server and let it sit idle or turn it off.

Even purchased on the used market, a decent server could cost $4000; if you spread that cost across 4 years, that's $80ish per month just to pay for the server, which is probably around your power / rent costs.

The costs are more stark if you're buying new, where it is easy to sped $10-30k for a decent spec server.

Don't buy a server to leave it turned off, you're wasting money.

Certainly for bit players and corporate datacenters with utilization < 1% you'd expect the median server to just sit there. For larger (amazon, google, etc) players the economic incentives against idleness are just too great.

> For larger (amazon, google, etc) players the economic incentives against idleness are just too great.

Not all workloads are CPU-bound. Cloud providers have many servers for which the CPUs are idle most of the time, because they're disk-bound, network-bound, other-server-bound, bursty, or similar. They're going to aim to minimize the idle time, but they can't eliminate it entirely given that they have customer-defined workloads.

But if the workload is not CPU-bound then why would they care about upgrading their CPUs to more performant ones, like Ice Lake Xeons?

The workloads are determined by their customers, and customers don't always pick the exact size system they need (or there isn't always an option for the exact size system they need). The major clouds are going to upgrade and offer faster CPUs as an option, people are going to use that option, and some of their workloads will end up idling the CPU. Major cloud vendors almost certainly have statistics for "here's how much idle time we have, so here's approximately how much we'd save with lower power consumption on idle".

If the workload is PCIe I/O bound, upgrading to PCIe Gen4 would be a huge benefit.

Reply on news.ycombinator.com