GHC eventlog based stats todo list

[harendra-kumar]

Changes to GHC, higher pri:

• Use perf counters in the user defined window as well
• Add a precise allocation counter
• Add a OS process level CPU counter as well to account for entire program time, this counter should not be stopped during the entire window. This would be a window only counter.
• flush the event log buffer before and after the thread run if beyond halfway mark

Changes to GHC, Lower pri:

• Add more relevant perf counters (esp. HW events)
• Add pinned allocations, unpinned allocations, and large allocations counters
• Enable the counter measurement code only when +RTS -l is specified (put code under eventlog_enabled)
• Put the code under ifdef LINUX_PERF_EVENT
• traceEventIO adds a lot of allocations, see if we can reduce that
• Use a unique id for each start/stop event pair for better correlation in case an event is missed
• Select the enabled counters via RTS options
• Support multiple counters at the same time
• Select a particular thread label to trace, to reduce eventlog traffic
• Select to log only the perf events

Changes to the eventlog parser:

• Print periodically e.g. every second instead of on each event
• Print after collecting n events
• activate a window across all threads, account the time of all haskell threads during the window