#129 -cyclic flag memory usage

[Daniel Kohlbek]

I'm trying to process data with the -cyclic flag but whenever I run it, my machine's memory gets used up resulting in the process being killed. Even when using the -U flag I have the same issue. Is it a very memory intensive process? I only have 16GB of ram on my machine and ~1GB of swap memory.

Here is the command I'm running:

dspsr -cont -K -d 4 -minram 750 -U 10240 -a psrfits -cyclic 1024 -E ../../../B0950+08.par -V ../../*.fits

Here is the last 20 or so lines of the -V output:

dsp::SingleThread::run SampleDelay done
dsp::SingleThread::run calling CyclicFold
dsp::Operation[CyclicFold]::operate
dsp::Transformation[CyclicFold]::operation call vchecks
dsp::Transformation[CyclicFold]::vchecks input checks
dsp::Transformation[CyclicFold]::vchecks done
dsp::Transformation[CyclicFold]::operation transformation
dsp::Fold::transformation
dsp::Fold::transformation choose_nbin
dsp::Fold::choose_nbin
dsp::Fold::choose_nbin folding_period=0.253088
dsp::Fold::choose_nbin sampling_period=0.000131072
dsp::Fold::choose_nbin minimum_bin_width=1.2 bins or 0.000157286 seconds
dsp::Fold::choose_nbin sensible nbin=1609
dsp::Fold::choose_nbin largest power of 2 < nbin=1024
dsp::Fold::choose_nbin using sensible nbin=1024
dsp::Fold::transformation call Fold::prepare_output
dsp::CyclicFold::prepare_output start
dsp::CyclicFold::prepare_output reset
dsp::Observation::copy_dimensions other->ndat=436166
dsp::TimeSeries::set_nbit (32) ignored
dsp::CyclicFold::prepare_output npol=4
dsp::PhaseSeries::resize nbin=1024
dsp::TimeSeries::resize (1024) data=0 buffer=0 ndat=436166
dsp::TimeSeries::resize reserve_ndat=0 fake_ndat=0
dsp::DataSeries::get_ndat_allocated  size=0 nval=4194304 bytes_per_dat=16777216 subsize=0 allocated_ndat=0
dsp::DataSeries::resize nsamp=1024 nbit=32 ndim=1 (current ndat=436166)
dsp::DataSeries::resize nbits=nsamp*nbit*ndim=32768
dsp::DataSeries::resize npol=4 nchan=1048576
dsp::DataSeries::resize nbytes=nbits/8*npol*nchan=17179869184 (current size=0)
dsp::DataSeries::resize set_ndat(1024)
dsp::DataSeries::resize require=17179869184
dsp::DataSeries::resize Memory::allocate (17179869184)
dsp::DataSeries::resize buffer=0x7fe561cb4010
dsp::DataSeries::resize memory_used=24327290880
dsp::DataSeries::resize reshape
dsp::DataSeries::reshape size=17179869184 bytes (subsize=4096 bytes)
dsp::DataSeries::resize return
dsp::PhaseSeries::resize_hits nchan=1048576 nbin=1024 hits_nchan=1
dsp::PhaseSeries::resize_hits Memory::do_allocate (4096)
PhaseSeries::zero this=0x558265594720
PhaseSeries::set_hits(0)
Killed

[Willem van Straten]

Hi Daniel, how many channels are in your input data? I'm guessing something around 1024, because the -cyclic 1024 later results in

dsp::DataSeries::resize npol=4 nchan=1048576

You can reduce the memory requirement by
* saving 2 polarizations instead of 4 (with -d 2) ... unless you have cyclic spectropolarimetry in mind
* reducing the number of phase bins from 1024 to 256 (with -b 256) ... very few pulsars have significant power beyond 100 harmonics (depending on the gain of the telescope)
* reducing the number of channels in the output periodic spectrum ... depends on the scintillation bandwidth that you're trying to resolve

[Daniel Kohlbek]

Hello, thank you for your suggestions. I think they resolved the memory issue, however I get the following error now:

Error::stack
        dsp::SingleThread::run
        dsp::SingleThread::run
        dsp::Operation::operate
        dsp::Transformation[CyclicFold]::operation
        dsp::Fold::transformation
        dsp::CyclicFoldEngine::fold
Error::InvalidState
Error::message
        Only Analytic input data is currently supported

Am I using the wrong format for the input data? For reference, my files are PSRFITS search mode data.

And yes, my input data is in 1024 channels.

Thank you again.

[Willem van Straten]

Hi Daniel, PSRFITS search mode data are detected (e.g. Stokes parameters or Coherency products) and to compute the cyclic spectrum requires undetected (e.g. baseband) data.

[Daniel Kohlbek]

Ah ok, sorry for the confusion. Thank you for the help!