Enabling perf_event support in runc

My team would like to contribute perf_event support for runc. We believe that taking limited amount of hardware counters available (e.g. 4 on Xeon 6240Y) a list of counters to be enabled should be passed to runc. It will allow users to fine-tune their counters and avoid multiplexing and scaling.

Therefore we suggest the following:

• extending config.Cgroup struct by adding PerfEvents []uint64 field
• extending cgroups.Stats struct by adding PerfStats map[uint64]PerfStat field
• defining PerfStat struct similar to kernel read_format struct
• using perf_event_open to start monitoring.
• enabling event grouping when multiple events are passed in config.Cgroup and not enabling if single event is passed

Our goal is to expose metrics collected using perf subsystem in cAdvisor and use them for advanced scheduling in clusters. Currently we use our own agent but customers' requirement is to switch to cAdvisor.

Maintainers, does it sound reasonable to you?

Review the config.Cgroup and cgroups.Stats definitions, then study the kernel perf_event_open API and its read_format structure. Confirm the proposed API with maintainers before implementation; done means configurable perf events, grouped handling for multiple events, returned statistics, and a path for cAdvisor to consume them.