MEMORY ALLOCATION ERROR! Compressed sensing (IRLS)

[Tomas Sikorsky]

Dear MDD comunity. I would like to ask for help with my problem. I'm running LINUX OS FEDORA 16 with PYTHON PYSIDE version 1.08 (I have tried 1.05 as well) and I get the following error message for CS IRLS algorithm.
IST algorithm on the other hand is extremely slow (about 100x slower than MDD with default parameters). I'm processing 3D NOESY with 15000 complex points of which 40% (6000) were recorded.

the error message from terminal:

####> mddnmr4pipeN 1 2 3 4 5

########## conversion from spectrometer format to nmrPipe format ../11 to ./XYZA/ft4sp.xyza.2D

Pipe streem (XYZA) ./XYZA/ft4sp.xyza.2D is ready

NLS pipe header ready ./XYZA/ft4sp.xyza

mddnmr4pipeN 1 2 3 4 5 failed

:

MEMORY ALLOCATION ERROR!

Thank you very much for your help

[Krzysztof Kazimierczuk]

Dear Tomas,

Could you please provide more details on your data. In particular:
NIMAX,
NI,
amount of RAM memory that you have,

also, could you please mention the number of iterations that you have set for IST.
Best Regards,
Krzysztof Kazimierczuk

W dniu 04.12.2012 19:47, Tomas Sikorsky pisze:

--
--
You received this message because you are subscribed to the Google Groups "mddnmr" group.
To post to this group, send email to mdd...@googlegroups.com rom your registered email address.
To unsubscribe from this group, send email to mddnmr-un...@googlegroups.com
For more options, visit this group at http://groups.google.com/group/mddnmr
 
 
 

-- 
Krzysztof Kazimierczuk, PhD
Laboratory of Intermolecular Interactions
University of Warsaw
Zwirki i Wigury 101, 02-089 Warsaw
tel. +48 22 8220211 ext. 421

[Tomas Sikorsky]

Thank you for your idea

I have 8GB of RAM. However the spectrum is quite big.
NIMAX parameters are:

NIMAX 160 96 1

The spectrum is recorded with 40%.

But if I reduce setenv NUS_POINTS below 5000 the IRLS works. Can you think of any trick how can I get the processing running with all 6000 points? I'm thinking about changing some parameter so that algorithm will be less memory demanding.

Thank you

P.S. for IST algorithm I have set 100 iterations

[Krzysztof Kazimierczuk]

Dear Tomas,

 There are several matrices allocated in the memory by IRLS algorithm, and largest of them is NIxNIMAXx16 bytes. The thing is, that NIMAX is in fact pre-processed for the need of algorithm:
1. It is rounded up to the nearest power of 2 (160 96) -> (256 128)
2. It is multiplied by CS_ZF parameter (256 128) -> (512 256)

This gives me about 6000x512x256x16=12.5 GB memory requirement.
With 5000 points it is ca. 10.4 GB. So I do not fully understand how it worked with 8 GB. Did it finish the processing?

You may try setting CS_ZF to 1 (default is 2) but this may affect the quality of reconstruction (but does not neccessarily have to - so try). This should shrink memory requirement 4 times.

Kind Regards,
Krzysztof

--
--
You received this message because you are subscribed to the Google Groups "mddnmr" group.
To post to this group, send email to mdd...@googlegroups.com rom your registered email address.
To unsubscribe from this group, send email to mddnmr-un...@googlegroups.com
For more options, visit this group at http://groups.google.com/group/mddnmr
 
 
 

[Tomas Sikorsky]

Thank you

I solved this problem by using computer with 16GB of RAM and all calculations finished correctly.

Regards Tomas Sikorsky


Dňa utorok, 4. decembra 2012 19:47:31 UTC+1 Tomas Sikorsky napísal(-a):