ConvertToMD performance

[Martyn Gigg]

This relates to #5682 and the comments surrounding the performance of the new ConvertToMD code.

I've attached a script and picture comparing the timings of running ConvertToMDEvents & ConvertToMD on the same input data. In this circumstance the code is faster. Is the performance degradation much more isolated than we think?

[Martyn Gigg]

Note: The MERLIN test data file is in the system test repository.

[Alex Buts]

The code degradation was occurring on large files (the real one) when multiple splitting was happening.

[Alex Buts]

refs #5710 this mainly fixes the issue.

The performance of the new code is not worth and even better (multithreading part) wrt the old code.

Changeset: 71fe48859f0a831713e7c22bf4ceda16ef869cc9

[Alex Buts]

the script testing performance for different parts of the code is provided

# Note that default conversion to events workspace compresses histogram data to have approximately 10 data-points per event. 
# The resulting MD workspace build from events contains ~40M events wrt the MD workspace build directly from histogram 
# containing ~184M events. The code converting from histogram in this example does much more work. 
#
Load(Filename='D:\Data\eclipse\workspaces\iliad\MER12524.raw',OutputWorkspace='MER12524',LoadMonitors=r'Separate')
SetUB(Workspace='MER12524',a='2.8700000000000001',b='2.8700000000000001',c='2.8700000000000001')
perfomance=1
ConvertToEventWorkspace(InputWorkspace='MER12524',OutputWorkspace='MER12524Ev')
##----------------------------------------------------------------------------------
## Old TMD Code in New Mantid; Adding to MD WS runs single threaded
ConvertToMDEvents(InputWorkspace='MER12524',OutputWorkspace='MER12524MD_old',QDimensions='Q3D',QConversionScales='HKL',dEAnalysisMode='Elastic',MinValues='-50,-50,-50',MaxValues='50,50,50')
#Duration 67 seconds;  In Mantid 2.1 was 77 sec
if perfomance:
	DeleteWorkspace('MER12524MD_old')

ConvertToMDEvents(InputWorkspace='MER12524Ev',OutputWorkspace='MER12524MDEv_old',QDimensions='Q3D',QConversionScales='HKL',dEAnalysisMode='Elastic',MinValues='-50,-50,-50',MaxValues='50,50,50')
# Duration 16 seconds;  In Mantid 2.1 was 16.79 
if perfomance:
	DeleteWorkspace('MER12524MDEv_old')

##----------------------------------------------------------------------------------
## Janik's 3D code
ConvertToDiffractionMDWorkspace(InputWorkspace='MER12524Ev',OutputWorkspace='CNS_MD',OutputDimensions='HKL')
#Duration 17 seconds
if perfomance:
	DeleteWorkspace('CNS_MD')
ConvertToDiffractionMDWorkspace(InputWorkspace='MER12524',OutputWorkspace='CNS_MD_HIST',OutputDimensions='HKL')
# Duration 56.76 seconds
if perfomance:
	DeleteWorkspace('CNS_MD_HIST')
##----------------------------------------------------------------------------------
## New MD code; Adding to MD ws runs multithreaded
ConvertToMD(InputWorkspace='MER12524',OutputWorkspace='MER12524MD_HistNew',QDimensions='Q3D',QConversionScales='HKL',dEAnalysisMode='Elastic',MinValues='-50,-50,-50',MaxValues='50,50,50')
#Duration 49 seconds TMD--  was 76.97 sec
if perfomance:
	DeleteWorkspace('MER12524MD_HistNew')

ConvertToMD(InputWorkspace='MER12524Ev',OutputWorkspace='MER12524MD_EvNew',QDimensions='Q3D',QConversionScales='HKL',dEAnalysisMode='Elastic',MinValues='-50,-50,-50',MaxValues='50,50,50')
# Duration 13 seconds
if perfomance:
	DeleteWorkspace('MER12524MD_EvNew')

[Alex Buts]

refs #5710 this mainly fixes the issue.

The performance of the new code is not worth and even better (multithreading part) wrt the old code.

Changeset: 71fe48859f0a831713e7c22bf4ceda16ef869cc9

[Roman Tolchenov]

Ran the script. The times of different algorithms looked comparable.

[Stuart Campbell]

This ticket has been transferred to github issue ​6556