Agenda:

 - Low latency CBC parameter estimation with spins with GPUs at Cascina

Attendees:

MAB, Stefano Cortese, Cristiano Palomba, Salvatore Vitale, Frederique Marion, CVDB, Livio Salconi, Riccardo Sturani, John Veitch, Elena Cuoco, Benoit Mours

Minutes:

John Veitch: several low latency PE developments. Different waveforms + techniques C+python code. No GPU interface. Spin is the issue. Latency of week with spinning waveforms (full spins). MPI MCMC / nested sampling code: not a large gain. Ben Aylott is investigating pyCUDA at Birmingham.

Riccardo: Ben Aylott is gaining a factor 10 wrt non GPU implementation. Key point: code in GPU friendly way, differential equations solver, which is a non parallel code. GPU allows parallelizing. But no spinning waveforms implementation on GPUs so far. Discussions with Leone Bosi about this: thinks this is possible.

John: maybe we could have a kind of hybrid implementation with GPUs for some part of the PE pipeline.

Gergely: what's steps are the time consuming steps. What steps are the bottle neck? template generation?

John: template generation in a nutshell. Filtering is OK. Sequential generation template. You generate one after the other.
Work by Ben Aylott: template in parallel.

Gergely: nb of waveforms to generate?

John 100k <--> 1M

Gergely: generation offline?
   
John: can't sample the parameter space

Gergerly: pre-calculated template bank?

John: the template bank could be generated. With resolution chosen in advance. Then, need to interpolate between template to have the needed resolution. That's not the present implementation.

Gergely: there would be advantages

John: it does not matter if you have to compute only once.

Gergely: what would be the length of the waveforms

John: size of the template bank is increasing with the low frequency. Millions of template.

Gergely: 1 M of 100s @ 4096 Hz (float)--> 2GB. Could be stored in GPU memory. It would be then possible to pre-calculate the template bank and load it on GPUs. Do filtering on GPUs and interpolate templates on GPU if needed.

Gergely: spinning waveform generation time?
Riccardo: less than a few s.
Gergely: Time of integration?
Riccardo : depends on the sampling rate.
Salvatore: 1kHz is enough for BBH. For BNS we need 4kHz.
Riccardo: integration step depends on the sampling rate.

Other subjects:
John: MBTA is very efficient, but this is still a coincident search. MBTA more coherent? Integration of parameter like inclination angle to get the sky localization.

Benoit: MBTA is coincident. Want to introduce coherence. Add coherence testing using best template in det1  applied in in det2.
Very low level of coherence.

Salvatore: use of MBTA best estimate as priori of PE code?

Frederique: there is a danger of bias, as MBTA estimation can be wrong.

Gergely: MBTA density of template 99% overlap. What is the PE MCMC MM?
John: there is no such MM. Resolution is defined by the code.
John: resolution is fixed by the errors on parameters. Template spacing is not defined a priori.
MAB: no uniform density.
John: yes, what could be done with a fixed template bank is to interpolate between templates. NR projects tried to do that as well.
Frederique: not a simple interpolation, but one needs to take into account.



Gergely: would be happy to learn about MCMC PE code.
John: will send you documentation/papers I wrote with Alberto Vecchio. But implementation on GPUs might be different.
Salvatore: injections 2 spins : ~ 1 week/injection. 50 injections TaylorF2 running: ready 

John: Leone was implementing a code on GPUs. Any news?

Gergely: TaylorF2 waveforms: done by other groups as well. Analytic waveforms. No in
Riccardo: discussions time domain generation of waveforms 6 months ago. But Leone has left Virgo since.

Gergely: the main pb is non analytic waveforms --> integration.

Salvatore: what we want to do, prompt emission EM or afterglow?
Frederique: prompt emission would mean minutes. So afterglow and actually the point is sky localization, to really point in the right direction.

Gergely: who should I contact for more info?
John: For what concerns MCMC/PE : Bayesian maiing. For more general 

MAB: Next steps: Salvatore results before the end of June.

Meeting adjourned