[PLUG] surprising performance on my HNOW nodes

[Russell Senior]

I've been doing some test runs on my heterogeneous network of
workstations and found some surprising things.  I've got 11 slave cpus
running as follows.  The percentage on the right is the fraction of
the job (10 million random walks) the associate CPU completed:
  
   (1x P4 2.53GHz)             14.89%
   (2x P3 1.0 GHz)             11.00% and 11.52%
   (1x P3 1.2 GHz)             16.18%
   (1x P3[Celeron] 1.0 GHz)    13.32%
   (1x P3 750 MHz)              9.98%
   (2x P3 500 MHz)              5.49% and 5.53%
   (1x P2 450 MHz)              5.01%
   (2x P2 300 MHz)              3.51% and 3.57%

The most surprising thing is how sucky the P4 performed.  The 1.2 GHz
P3 did better, and it was even farther away on the network (about 40
ms latency over a 256kbps link, as opposed to 0.1 ms latency on a 100
Mbit ethernet).  Neither box had any significant load on them.

This is an almost exclusively CPU bound workload.  Hardly any memory
consumption (RES ~ 1370K), no disk I/O and very little network I/O.
Mostly floating point operations.  The same binary is running on all
nodes, all working on the identical problem.

The other slightly surprising thing is that the 1GHz Celeron did
better (per CPU) that the P3 1GHz, but the latter is my workstation
where all this crap is scrolling away, so it is possible to
rationalize that one.

Any clues about the P4 stinkiness?

-- 
Russell Senior         ``I have nine fingers; you have ten.''
seniorr at aracnet.com