----- Original Message -----
>
> nathans wrote:
>
> > [...] Thinking system wide filters like "when looking at
> > per-process metrics, this system is expected to have N million
> > insts, so force a system wide filter for all access to per-process
> > metrics via a specified cgroup". [...]
>
> Would this be a administrative preference for security/load-limiting?
Hmm, more like "for all processes that should appear in daily pcp logs
or appear in a pmatop listing, they need to belong to cgroup X" - and,
mechanisms for getting arbitrary processes in/out of said cgroup X are
made readily accessible for people doing analysis.
> If so, it's bound to be ineffective.
For some people it will be ineffective, for others they wont care (but
sure will when their root disk fills because we auto-logged too much),
and for yet others it will be highly effective. This is the same sort
of approach the existing perf_event cgroup uses, FWIW, & I imagine for
some its highly useful, others wont care and yet others are happy with
system-wide perf events. Its better to have the option than no way to
deal with this class of problem though.
cheers.
--
Nathan
|