Re: [boost] [Block Pointer] benchmark

Subject: Re: [boost] [Block Pointer] benchmark
From: Bryce Lelbach (blelbach_at_[hidden])
Date: 2011-05-26 00:25:51

On 2011.05.24_16.52.00, Phil Bouchard wrote:
> FYI make_shared is slower than new because of the temporary it
> creates.  If people what speed they should stick to operator new in
> all cases.

My box begs to differ.

Units for min, max, mean and moment are: shared_ptrs created / 8 seconds
Count is: number of trials
Sum is: shared_ptrs created / (numer of trials * 8 seconds)

GCC 4.4.5:

  make_shared:
    min       = 38176474
    max       = 40958919
    mean      = 3.95662e+07
    count     = 8
    sum       = 316529282
    moment<2> = 1.56663e+15
  new:
    min       = 32075900
    max       = 35917824
    mean      = 3.41757e+07
    count     = 8
    sum       = 273405391
    moment<2> = 1.16916e+15

Intel C++ 12.0 

  make_shared:
    min       = 38264052
    max       = 40850617
    mean      = 4.0022e+07
    count     = 8
    sum       = 320176020
    moment<2> = 1.6026e+15
  new:
    min       = 33708800
    max       = 35364869
    mean      = 3.47297e+07
    count     = 8
    sum       = 277837391
    moment<2> = 1.20645e+15

Clang 2.8

  make_shared:
    min       = 35269466
    max       = 37759937
    mean      = 3.69983e+07
    count     = 8
    sum       = 295986798
    moment<2> = 1.3695e+15
  new:
    min       = 32876828
    max       = 34609212
    mean      = 3.38261e+07
    count     = 8
    sum       = 270608628
    moment<2> = 1.14456e+15

lll-clang 3.0

  make_shared:
    min       = 34756169
    max       = 37310038
    mean      = 3.66762e+07
    count     = 8
    sum       = 293409533
    moment<2> = 1.34575e+15
  new:
    min       = 33098433
    max       = 34011625
    mean      = 3.37868e+07
    count     = 8
    sum       = 270294200
    moment<2> = 1.14162e+15

Equipment and kernel:

  HP ProLiant DL785 G6 Server
    8 AMD Socket Fs, 6 Opteron 8431s per socket w/ HyperTransport
      Cache sizes per socket: L1 384kB, L2 3072kB, L3 5120kB 
    96 gigs main memory, 100ish gigs swap

  RHEL 6.0 (Linux 2.6.32), sits at runlevel 3

  Boost from top of trunk

No other jobs were running at the time.

Peak memory usage for one invocation of the benchmark program is appx 60 gigs.

Source code attached.

-- 
Bryce Lelbach aka wash
boost-spirit.com
px.cct.lsu.edu
github.com/lll-project