Understanding the Performance and Power Saving Tradeoffs of Server Sleep States

With the growing use of software based networking functions run on servers and the increased pressure to save power, it is more important than ever for telecommunications companies to fully leverage the power saving tools available on this form of hardware. It is often the case that servers spend a disproportionate amount of time in an idle state; therefore, we investigate a range of server settings which enable idle servers to enter sleep states. Although these settings have been accessible for some time, they are generally underutilised. We run tests on Intel architecture (IA) x86 servers to measure the potential power savings of different sleep states, considering this against the cost in exit latency. We present a comprehensive breakdown of the pros and cons of each sleep state, ultimately allowing them to be maximally utilised in a telco environment with confidence. We find that the maximum additional power saving of all the considered sleep states comes from core C-states, with a maximum ~80% processor power saving when idle. Package C-states realise an additional power saving but comes at the cost of an increased exit latency.