README.performance

Written by:
  Kevin J. Bowers, Ph.D.
  Plasma Physics Group (X-1)
  Applied Physics Division
  Los Alamos National Lab
April 2004 - original version

7.8 million particles advances per second per processor sustained in the
common case (ss2.lanl.gov cluster; 2.533 GHz Pentium 4 w/ Gigabit Ethernet
v4 hardware acceleration enabled).

This corresponds to ~1.92 Gflop/s sustained per node in the particle advance
(~76% of theoretical) and ~1.74 GB/s memory movement to/from floating point
unit sustained. The basis for the Gflop/s and GB/s estimates are below:

To the author's knowledge, this is the fastest, most efficient, fully
relativistic 3d particle in cell simulation in the world.

---------------------

Per particle, the common case advance requires approximately:

60 add
80 multiply
1  rcp
2  rsqrt
6  compares
20 SIMD bit ops (logicals, conditional moves, conditional zeros)

1 rcp   ~ Initial approximation cost + 2 add + 2 multiply
1 rsqrt ~ Initial approximation cost + 2 add + 4 multiply

Taking initial approximation cost of the rcp and rsqrt to be four times the
polishing cost and assuming an add, multiply, compare and vector bit op all
cost one flop yields:

246 flops per particle in common case

For memory traffic, per particle, the common case requires:

32 bytes memory loads (particle)
80 bytes L2 loads (interpolator)
48 bytes L2 loads (accumulator)
32 bytes memory stores (particle)
48 bytes L2 stores (accumulator)

This corresponds to 240 bytes shuffled in/out of the floating point unit per
particle (160 incoming, 80 outgoing).