perflog.txt

[2011.02.18] 
A log for verbose notes pertaining to performance.
--------------------------------------------------



[2011.02.18] {Initial blackscholes runs}

First a version of blackscholes that uses the parMap in par.hs.

GHC 7.0.1, released version.
Listing a couple runs at each setting.

Running on a 4-core 3.33 ghz nehalem, SMT on (with -qa)
  10M options with 10K blocksize:
    1 thread -- 5.1 s
    2        -- 6.25s  5.87s real/user
    2 thr #2 -- 3.8s   6.4s 
    3        -- 2.8s   5.38s
    3        -- 2.0s   5.86s
    4        -- 1.8s   5.4s 
    4        -- 1.3s   5.15s 
    5        -- 1.5s   5.4s 
    5        -- 1.15s  5.6s 

High variance... definitely thrashing sometimes.  It very rarely uses
over 300% cpu at 4 threads.

Removing the "-qa" option doesn't seem to change things much.
Even when it does badly its only spending 1% of time in GC.

  Next let's try AList's parBuild.
  --------------------------------

Not working right yet...


What about mandel?
  It's getting <2X speedup for me on four threads on this machine atm.
  It's about the same with the factored AList/parBuild as opposed to
  the original.