Implement New Time Tracking

[ehuss]

Using TSC has a variety of problems.

• Determining which sysctl to read to determine the TSC-to-seconds conversion is difficult. Some systems don't even have it (My Mac Mini with Snow Leopard, Bob's Amazon FreeBSD Xen virtualized system, etc.).
• Some systems do not have a reliable TSC. We do not check for this.
• There's clunky code to handle re-calibrating our relationship of TSC to wall clock. I've never been comfortable with this. (This was a performance hack to avoid calling gettimeofday a lot, I'm not sure how important that is.)

We need a new API for handling high-resolution time. Preferably something that can use multiple types of time. See https://github.com/ironport/aplib/issues/2 for more detail.

Preferably, this new API would be in shrapnel to help remove the dependency with aplib.

[ehuss]

BTW, I found a really interesting document that talks about timing and TSC issues: http://www.mjmwired.net/kernel/Documentation/virtual/kvm/timekeeping.txt

[samrushing]

We probably want clock_gettime (MONOTONIC_FAST) where we can get it.

Not sure yet about Darwin, since mach_absolute_time() is returning a scaled version of TSC on my machine (i.e., I have a 2.26GHz machine and it is scaled by exactly that amount to return nanoseconds since boot).

I would like for the TSC code to stay in there as an option for those platforms that don't care about the various issues. Another fun thing might be to implement RDTSCP and keep track of TSC on each CPU?

[samrushing]

Another thing that might be nice: integrating the SNTP code. Did that make it out?