Fixing opal_timer_linux_find_freq() on x86

[zzzoom]

Due to the TSC frequency debacle in #3003 MPI_Wtime got switched back to clock_gettime() but opal_timer_linux_find_freq has been left broken. This only affects the one remaining call to opal_timer_base_get_freq left in opal_progress_set_event_poll_rate but I'd like to take a shot at fixing it anyway.

Some alternatives I've been thinking about:

1. Measure the frequency between a couple of samples of rdtsc and clock_gettime. The longer the period the better, but opal_timer_linux_find_freq doesn't do much so it would take calling usleep for a few µs or taking the measurements elsewhere during opal_init. There's a small difference between sockets, so context switches + migration would be problematic. Ugh.
2. Take bogomips instead of cpu MHz from /proc/cpuinfo, divide by 2 and hope it keeps working in the future. Linux tries to avoid breaking userspace, so it's probably a safe bet.
3. Make it a configurable parameter and let users/admins worry about it.

Any comments?

[bosilca]

According to the discussion on the MPI Forum #77 modern Intel processors have rdtsc ticking at the nominal processor frequency, regardless of power saving and frequency boosting (I paraphrased Jim Dinan).

[jdinan]

That's correct -- You can check for invariant TSC to confirm that the processor supports and is configured to provide this behavior. See: https://software.intel.com/en-us/forums/intel-isa-extensions/topic/280440. Linux should also set the constant_tsc flag in /proc/cpuinfo.