PmcResearch [FreeBSD WIKI]

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== Pmctools: Research ======

This page lists PMC related research that could be done.

===== A GUI Visualizer =====

This tool helps a user to visually 'see' hotspots in the code and
correlate these to source code.  The tool will need to:

  * Attach process-mode PMCs to selected interrupt threads, or processes and their descendants.
  * Sample the whole system or specific threads/processes.
  * Show the progression of 'hot-spots' in a process or in the kernel over time (i.e., a 3-d graph).
  * Show cache miss / tlb miss / instruction retiral / CPI data of a process over time (i.e., a 2-d graph).  Aggregate this data for all processes corresponding to a given executable.
  * Show source code corresponding to selected hot-spots in a pane (or show a disassembly listing if we can't locate sources).
  * Profile a remote target system (needs some kind of controlling daemon on the target).

The 'research' component here lies in finding neat ways of displaying
collected performance data to the user.

==== Related Reading ====
  * [[http://worrydream.com/MagicInk/|Magic Ink Information Software and the Graphical Interface]], by Bret Victor.

===== Performance speed ups =====

Use the data in hwpmc(4) logs to intelligently reorder executables.  For example, we could:
  * Group 'hot' functions into contiguous virtual addresses to minimize the usage of physical memory.
  * Order functions to improve startup time for executables.

We could have 'soft' PMCs (see [[PmcTools/PmcKinds|PmcTools/PmcKinds]]) that 'sample' on page
faults, in addition to samples from true PMCs.

===== Drive 'cc -pg' profiles from PMC hardware =====

This is an extension of existing profiling infrastructure.  In it we
integrate conventional profiling (timer based profiling) with a 'clock'
derived from the PMC sampling interrupt.  We'd change ''%%monitor()%%'' to
allocate a PMC if a specific environment variable is present.

===== PMCs and Scheduling =====

Use PMC-based measurements to drive scheduling decisions.  For example,
measuring cache activity at the end of a CPU slice could help us measure
how 'cold' the cache has become w.r.t., other processes.

===== PMC driven static analysis of code =====

Use static analysis and data from CPU cache behaviour to identify
areas of code where we could reduce cache thrashing and improve
cache locality.