
[Science] [BAS home] [Met home] [Beowulf home]	Antarctic Meteorology

AH's timings (needs password) | AI's timings

Timings

All runs have about the same stash set: 30 numbers, about 100 items. See the indiviual job specs.

Except where noted these timings are on the "minifer" cluster (Athlon 1600+ MP's).

Contrary to what I understand to be other peoples experience, dual-boards run as fast as two singles.

Timer output for 1x1 "MPP" coupled 64-bit jobs sans meaning

Opteron

I have the coupled model running on the opteron cluster, but it drifts. Nonetheless the timings may be of interest: about 1.2 years/day. See the cpuinfo if you want.

Summary

To go here...

Atmos Only

64-bit (yabcc)

Procs		Mins/day		Speedup		Efficiency		ModelYears/RealDay
1x1		    			-		1.0
1x2		4.217 (y1) 							About 0.95
                4.236 (y2)
                4.225 (y3)

On a sunblade-2 (980 MHz) I get 6.4 mins/day. See sun for details.

32-bit (yabcc)

Procs		Mins/day		Speedup		Efficiency		ModelYears/RealDay
1x1		    			-		1.0
1x2		2.253 (y1)							About 1.8
		2.258 (y2)

On a P4 2GHz I get 5.23 mins/day on 1 proc.

Coupled

64-bit

Procs           Mins/day                Speedup         Efficiency              ModelYears/RealDay
1x2             8.00000                 -               1.0                     0.50

32-bit

Procs           Mins/day                Speedup         Efficiency              ModelYears/RealDay
1x1             7.84444                 -               1.0                     0.51

The stuff below here is now a bit aged...

Atmos Only

64-bit

(on the old 1G cluster)

Procs		Mins/day		Speedup		Efficiency
1x1		13.3			-		1.0
1x2		 6.929 			1.919		0.960
1x3              5.733			2.320		0.773			[note: over only 15 model days]
1x4		 5.345			2.489		0.622			[note: only based on 449 ts]

On RCAHs bslpsdf (yaapf; expected to be 1.7x as fast as master based purely on MIPS):

Procs		Mins/day		Speedup		Efficiency	Speedup-over-master
1x1		 7.239			-		1.0		1.84	[Note: only 30 model days]
1x2		 4.4			1.645		0.822		1.57 	[Note: only 15 model days]

On my 2GhZ Piv (yaaug; some sub-optimisations for bit-cf):

Procs           Mins/day                Speedup         Efficiency      Speedup-over-master
1x1		 8.0                    -               1.0

Something wrong here: slower than Richards 1.7G! Probably because its a slower job.

32-bit

(on the old 1G cluster)

Procs		Mins/day		Speedup		Efficiency
1x1		6.981 			-		1.0
1x2		3.556			1.963		0.982			[Note: only 2 months]

On RCAHs bslpsdf (yaaph)

Procs           Mins/day                Speedup         Efficiency      Speedup-over-master
1x1              4.633                  -               1.0             1.501	[Note: 30 days]
1x2              2.666                  1.737           0.869           1.34	[Note: 30 days]

Ocean Only

64-bit

yaasa. 98x73x20. This isn't what we want to run...

Procs           Mins/day                Speedup         Efficiency     
1x1              1.2			-		-			[Note: 20 days]
1x2		 0.6			2.0		1.0			[Note: 20 days]

Coupled

64-bit

yaasc. "hadcm3".

Procs           Mins/day                Speedup         Efficiency
1x1             26.7333                 -               -                      
1x3             11.348                  2.36            0.785                    [Note: 24 days]

On RCAH:

Procs           Mins/day                Speedup         Efficiency
1x1             16.1333                 -               -
1x2		 9.9			1.630		0.815			[Note: 1 days]

32-bit

yaase. Stash???

Procs           Mins/day                Speedup         Efficiency
1x1             14.00                   -               -                      
1x2              7.72                   1.81            0.907                    [Note: 2 months]

On RCAH (yaase):

Procs           Mins/day                Speedup         Efficiency
1x1             10.9                    -               -                      	 [Note: 1 month]
1x2              6.57                   1.66            0.830                    [Note: 1 month]

CPU info

As in "more /proc/cpuinfo".

master: master-cpu [997 Mhz; 1986 bogomips]
bslpsdf: bslpsdf-cpu [1694 Mhz; 3382 bogomips]. Richards machine is a dual processor board. We were told that dual-procs don't work well. But it seems fine. Odd.
my new [2002/02/14] Piv 2GHz: p4-2000-cpu
minifer: minifer-cpu [1393.801 Mhz; 2778.72]. Athlon 1600 MP's. Note that although the clock is lower, the model runs *faster* than on the P4.

In fact bogomips and CPU MHZ seem to be in ratio:

	1694/997. =  1.69910
	3382/1986. = 1.70292

Past last modified: 1/10/2004 / wmc@bas.ac.uk