sfgen Benchmark - Scaling Studies

Fixed Passes

Rather than burn cycles, the number of passes will be fixed at 16, for all processor counts. This should sufficient to reduce the noise in each pass, and will provide a reasonable number for extrapolating the required time for each job.

1024³

These measurements were made in two batches.

Small Bins

Parameter	Value
NumberOfProcessors	32
NumberOfPasses	128
NumberOfBins	8
NumberOfPDFBins	5000
NumberOfPairs	1048576
StartLength	0.00390625
EndLength	0.015625

Large Bins

Parameter	Value
NumberOfProcessors	32
NumberOfPasses	128
NumberOfBins	5
NumberOfPDFBins	5000
NumberOfPairs	524288
StartLength	0.022097086912079612
EndLength	0.25

Scaling Parameters

For each bin, the total number of points is NumberOfPairs x NumberOfProcessors x NumberOfPasses. For each bin, we want to hold NumberOfProcessors x NumberOfPasses fixed, at 4096. This gives the following:

NumberOfProcessors	NumberOfPasses
32	128
64	64
128	32
256	16
512	8
1024	4

Starting at 32 is to make sure the data can be loaded into memory for all tests.

2048³

Parameters used on BigBen for 2048³ simulation.

Parameter	Value
NumberOfProcessors	4096
NumberOfPasses	64
NumberOfBins	13
NumberOfPDFBins	5000
NumberOfPairs	65536
StartLength	0.0078125
EndLength	0.5

Scaling Parameters

NumberOfProcessors x NumberOfPasses = 262144. This gives the following:

NumberOfProcessors	NumberOfPasses
32	8192
64	4096
128	2048
256	1024
512	512
1024	256

I'll need to check to find out where the cutoff is for memory. 32 tasks is pretty unlikely.