# First round of timing results
# Single processor on 300 MHz master ganesh, no PVM
0.880user 21.220sys 99.9%, 0ib 0ob 0tx 0da 0to 0swp 0:22.11
0.280user 21.760sys 100.0%, 0ib 0ob 0tx 0da 0to 0swp 0:22.04
# Single processor on 400 MHz slave b4 using PVM
0.540user 11.280sys 31.3%, 0ib 0ob 0tx 0da 0to 0swp 0:37.65
0.700user 11.010sys 31.1%, 0ib 0ob 0tx 0da 0to 0swp 0:37.62
# 2x400 MHz (b4, b9) with PVM
1.390user 14.530sys 38.3%, 0ib 0ob 0tx 0da 0to 0swp 0:41.48
# 3x400 MHz (b4, b9, b11) with PVM
1.800user 18.050sys 46.5%, 0ib 0ob 0tx 0da 0to 0swp 0:42.60

# OK, this is terrible.  We make two changes.  First, we eliminate
# printout statements in slaves (which were logging to a file).  There
# are a LOT of calls -- this adds significant overhead and corrupts
# the timing runs.  Second, we no longer send slices to each host --
# we just send the slice structure along with the weights and await
# the results.  Fewer calls, and above all, ONLY multicasts from the
# master.  The only REAL question is whether we want to deliberately
# desynchronize the slaves by a tiny bit to keep their returns from
# colliding!  We'll do the single processor again because it might
# have changed...
#
# Single processor on 300 MHz master ganesh, no PVM.  Guess not.
1.250user 20.630sys 99.9%, 0ib 0ob 0tx 0da 0to 0swp 0:21.90
# Single processor on 400 MHz slave b4 using PVM.  Better.
0.350user 10.460sys 32.9%, 0ib 0ob 0tx 0da 0to 0swp 0:32.79
2.380user 8.410sys 32.5%, 0ib 0ob 0tx 0da 0to 0swp 0:33.11
# 2x400 MHz (b4, b9) with PVM
2.260user 11.140sys 37.7%, 0ib 0ob 0tx 0da 0to 0swp 0:35.53
# 3x400 MHz (b4, b9, b11) with PVM
1.630user 11.160sys 40.3%, 0ib 0ob 0tx 0da 0to 0swp 0:31.67
# 4x400 MHz (b4, b9, b11, b12) with PVM
2.720user 14.720sys 42.9%, 0ib 0ob 0tx 0da 0to 0swp 0:40.61

# Still no gain, but close.  We try increasing the granularity a
# bit by using a bigger dataset.
#
# Single processor on 300 MHz master ganesh, no PVM.  Takes longer.
9.270user 207.020sys 99.9%, 0ib 0ob 0tx 0da 0to 0swp 3:36.32
# Single processor on 400 MHz slave b4 using PVM.  Better.
4.380user 61.410sys 28.3%, 0ib 0ob 0tx 0da 0to 0swp 3:51.67
# 2x400 MHz (b4, b9) with PVM.  At last a distinct benefit!
3.080user 71.420sys 51.1%, 0ib 0ob 0tx 0da 0to 0swp 2:25.73
# 3x400 MHz (b4, b9, b11) with PVM. Still better.
1.270user 70.570sys 58.9%, 0ib 0ob 0tx 0da 0to 0swp 2:01.89
# 4x400 MHz (b4, b9, b11, b12) with PVM. And peak.
6.000user 71.820sys 63.3%, 0ib 0ob 0tx 0da 0to 0swp 2:02.83
# More processors would actually cost speedup at this granularity.