Message passing benchmark on smp emulator

[Vance Shipley]

I ran some tests on an UltraSparcT1 using P11B.  This is a
T2000 server with 8 1.0 GHz cores and 16GB of memory.

To get the build to work I had to patch the configure
script to recognize the architecture:

  ARCH=noarch
  case `uname -m` in
  sun4u)        ARCH=ultrasparc;;
+ sun4v)        ARCH=ultrasparc;;
  i86pc)        ARCH=x86;;
  i386) ARCH=x86;;
  i486) ARCH=x86;;


Using Rickard Green's big:bang/1 benchmark I got the following numbers
(in seconds) for various numbers of schedulers:

            50       100      200      300       400       500       600       700        800
--------- -------- -------- -------- -------- --------- --------- --------- --------- ----------
R10B      0.043230 0.277980 1.833356 5.600693 14.786129 30.843167 54.963973 86.961122 133.321735
P11B (1)  0.041275 0.232576 1.396134 4.262447  9.660085 18.410988 31.857533 56.252369  90.940019
P11B (2)  0.023220 0.125535 0.752071 2.263410  5.571316 11.274116 20.548645 36.046551  54.599152
P11B (4)  0.014485 0.074506 0.420364 1.280111  3.298273  6.505116 11.829685 20.762503  31.040196
P11B (8)  0.011881 0.055231 0.279430 0.822115  2.094020  3.947909  7.083040 11.900371  17.554864
P11B (12) 0.011408 0.055218 0.269419 0.736126  1.725916  3.298313  5.897804  9.582168  13.901548
P11B (16) 0.012796 0.058613 0.264527 0.715797  1.611627  3.023646  5.279931  8.100685  11.655361
P11B (20) 0.013608 0.060831 0.272990 0.708398  1.524395  2.940332  4.953840  7.301772   9.937516
P11B (24) 0.013591 0.066613 0.287982 0.704209  1.506340  2.820041  4.519642  6.596950   9.603010
P11B (32) 0.016180 0.070481 0.305068 0.763517  1.541233  2.729289  4.480478  6.492539  10.198520


The average decrease in execution time for the number of schedulers:

  2      4      8      12     16     20     24     32
------ ------ ------ ------ ------ ------ ------ ------
41.70% 66.12% 78.04% 80.60% 81.24% 81.50% 81.54% 80.35%


So in summary the port to the UltraSparcT1 was trivial and there
was an immediate benefit.  This test, using this snapshot, however
does show diminishing returns when the number of schedulers gets 
above four.

	-Vance


On Tue, Mar 07, 2006 at 05:52:26PM +0100, Rickard Green wrote:
}  
}  In order to be able to take advantage of an smp emulator I wrote another
}  message passing benchmark. In this benchmark all participating processes
}  sends a message to all processes and waits for replies on the sent messages.