Forked from
CNAF / Annual Report / Annual Report 2018
434 commits behind the upstream repository.
AdV_computing_CNAF.tex 14.73 KiB
\documentclass[a4paper]{jpconf}
\usepackage{graphicx}
\begin{document}
\title{Advanced Virgo computing at CNAF}
%\author{P. Astone$^1$, F. Badaracco$^{2,3}$, S. Bagnasco$^4$, S. Caudill$^5$, F. Carbognani$^6$, A. Cirone$^{7,8}$, G. Fronz\'e$^{4}$, J. Harms$^{2,3}$, I. LaRosa$^1$, C. Lazzaro$^9$, P. Leaci$^1$, S. Lusso$^4$, C. Palomba$^1$, R. DePietri$^{11,12}$, M. Punturo$^{10}$, L. Rei$^8$, L. Salconi$^6$, S. Vallero$^{4}$, on behalf of the Virgo collaboration}
\author{P. Astone$^1$, F. Badaracco$^{2,3}$, S. Bagnasco$^4$, S. Caudill$^5$, F. Carbognani$^6$, A. Cirone$^{7,8}$, M. Drago$^{2,3}$, G. Fronz\'e$^{4}$, J. Harms$^{2,3}$, I. LaRosa$^1$, C. Lazzaro$^9$, P. Leaci$^1$, S. Lusso$^4$, C. Palomba$^1$, R. DePietri$^{11,12}$, M. Punturo$^{10}$, L. Rei$^8$, L. Salconi$^6$, S. Vallero$^{4}$, on behalf of the Virgo collaboration}
\address{$^1$ INFN, Roma, IT}
\address{$^2$ Gran Sasso Science Institute (GSSI), IT}
\address{$^3$ INFN, Laboratori Nazionali del Gran Sasso, IT}
\address{$^4$ INFN, Torino, IT}
\address{$^5$ Nikhef, Science Park, NL}
\address{$^6$ EGO-European Gravitational Observatory, Cascina, Pisa, IT}
\address{$^7$ Universit\`a degli Studi di Genova, IT}
\address{$^8$ INFN, Genova, IT}
\address{$^9$ INFN, Padova, IT}
\address{$^{10}$ INFN, Perugia, IT}
\address{$^{11}$ Universit\`a degli Studi di Parma, IT}
\address{$^{12}$ INFN, Gruppo Collegato Parma, IT}
%\address{Production Editor, \jpcs, \iopp, Dirac House, Temple Back, Bristol BS1~6BE, UK}
\ead{luca.rei@ge.infn.it}
\begin{abstract}
Advanced Virgo (AdV) is a gravitational wave (GW) interferometric detector located near Pisa, Italy. It's a Michelson laser interferometer with 3 km long Fabry-P\'erot cavities in both arms; it is the largest GW detector in Europe and it operates in network with the two LIGO detectors in the US. Together we are part of the LIGO-Virgo Collaboration (\emph{LVC}). During the first two observing runs (O1 and O2), the LVC was able to detect more than 15 possible GW signals, among which 11 confirmed and 1 with an electromagnetic counterpart. LVC has just started a new observing run (O3) on April 1, with a planned duration of 12 months. The spectral sensitivity of the three detectors has largely increased over the last year and with it also the accessible volume of space, leading to an event rate of roughly one per week for Binary Black Hole (\emph{BBH}) mergers and one per month for coalescing Binary Neutron Stars (\emph{BNS}). With that in mind, the LVC is developing new cutting-edge data analysis pipelines, also to identify and study related electromagnetic (\emph{EM}) counterparts with a very low-latency from the GW events. Starting from last year both GW data and software are freely released to the EM follow-up partners to let them support our analysis. At the same time the Gamma-Ray Coordinates Network (\emph{GCN}) will automatically trigger and coordinate all the telescopes of the collaboration and release public alerts. The era of Multimessenger astronomy has just begun and CNAF is going to play a key role for the Virgo collaboration data management.
\end{abstract}
\section{Advanced Virgo 2018-2019 achievements}
%The analysys of data aquired in the 2018 has enlightened some aspect of general relativity, proving that our models are quiete well-founded, and explored other gravity theoryes. To explore better the science of gravity the LVC has worked in the last year to greatly increased the sensibility of all three interferometers and improved pipeline analysis and GW algorithmics. Virgo is now able to sense a `standard' BNS up to 50 Mpc, almost doubling the BNS range during O2 of nearly 27 Mpc. Virgo data acquisition system scaled conseguently, passing from 36MB/s to ~ 50-60MB/s in the commissioning phase and during scientific run (after removing some channels) is narrowed to ~ 35MB/s. All this leads to a partial reorganization of data transfer and management and on the way virgo computes.
The amount of data processed during the last few years has emphasized the fact that our General Relativity based models are considerably robust, while still leaving some room for alternative modified gravity theories. In order to investigate further, LVC is working hard to improve the detector performances and expand the sensible Universe horizon, which for instance is now up to 50 Mpc for AdV and for BNS merger events. The AdV data acquisition system ha scaled in parallel, moving from 36 MB/s to 50-60 MB/s during the commissioning phases and stabilizing at 35 MB/s during the scientific run (O3). This has been achieved thanks to a partial reorganization of the data transfer, management and computing facilities.
\section{Advanced Virgo computing model}
\subsection{Data production and data transfer}
The Advanced Virgo data acquisition system is writing about 35MB/s of data (so-called ``bulk data'') during O3. CNAF and CC-IN2P3 are the Virgo Tier-0: during the science runs, bulk data is stored in a circular buffer located at the Virgo site, and simultaneously transferred to the remote computing centres where they are archived in tape libraries. The transfer is realized through an ad-hoc procedure based on GridFTP (at CNAF) and iRods (at CC-IN2P3). Other data fluxes reach CNAF during science runs:
\begin{itemize}
\item trend data (few GB/day), periodically transferred using the system described above;
\item Virgo-RDS or Reduced Data Set (about 100GB/day), containing the main Virgo channels including the calibrated dark fringe. This set of data is currently transferred from Virgo to the LIGO computing repositories using LDR (LIGO Data Replicator), but plans are to use Rucio instead shortly (while still using iRODS to CCIN2P3);
\item LIGO-RDS, containing the reduced set of data produced by the two LIGO detectors and analysed at CNAF, transferred through LDR;
\end{itemize}
\subsection{Data Analysis at CNAF}
%The analysis of the LIGO and Virgo data was made jointly by the two collaborations; the analysis pipelines are distributed among the worldwide network of computing facilities offering computing resources to the GW experiments. CNAF was mainly used for CW analysis, looking for continuous gravitational wave signals, developed by INFN–Roma people (see hereafter more details). But at CNAF is also running part of the pyCBC pipeline, submitted via OSG, looking for compact binaries signals. pyCBC has a crucial role in the detection of the coalescence of BBH and BNS. CNAF contributed to the computation performed through pyCBC for the analysis of the events GW170814, the first BBH coalescence detected also by Virgo, and GW170817, the BNS coalescence. During the last month a new extension of CVMFS, \emph{big cvmfs} was mounted at cnaf to support another OSG pipeline, \emph{BayesWave}. The big cvmfs is able to export, in a posix fashion, big file of data from nearby cache in Amsterdam instead of accessing data directly from Nebraska. BayesWave is a Bayesian algorithm designed to robustly distinguish gravitational wave signals from noise and instrumental glitches without relying on any prior assumptions of waveform morphology. In the last year coherent WaveBurst \emph{cwb} was ported to cnaf and made available to run. cwb is a pipeline based on coherent algorithm for detection and reconstruction of modelled and unmodelled GW bursts. A new newtonian noise cancellation algoritmh, developed by the group of Gran Sasso Science Institute (\emph{GSSI}) was made available very recently. The increased number of LVC pipelines running at cnaf has led to saturate advance virgo pledge at cnaf, cnaf promptly rensponded to advance virgo needed enlargin our quota and giving experimental access to gpu.
LIGO-Virgo data analysis is organized jointly, meaning that the analysis pipelines are made available to the computing facilities related to the LVC network, ready to be distributed to each GW detector. CNAF has been mainly used for Continuous Wave(\emph{CW}) analysis, led by the Roma INFN group, and for the Compact Binary Coalescence python-based analysis (\emph{pyCBC}), submitted via OSG. In particular CNAF computationally contributed to GW170814 and GW170817 events, respectively the first BBH coalescence detected by Virgo and the first BNS merger ever observed. During the last month a new extension of CVMFS, so-called ``big cvmfs'', was mounted at CNAF to support another OSG-based pipeline, Bayes Wave. The former is able to make available, in a POSIX-like fashion, big data files from a cache in Amsterdam, instead of accessing the data directly from Nebraska. The latter is a Bayesian algorithm, designed to robustly distinguish GW signals from noise and instrumental glitches, without relying on any prior assumptions on the waveform shape. During the last year, coherent WaveBurst(\emph{cWB}), an algorithm dedicated to the detection and reconstruction of GW Bursts, was also ported to CNAF. Furthermore, new Newtonian Noise cancellation algorithms, which are currently being developed by the GSSI group, were made recently available. The increasing number of LVC pipelines running at CNAF has led to resource saturation, and consequently to a demand for enlarged computing power, together with access to GPUs.
\subsubsection{CW pipeline}
CNAF has been in 2018 the main computing center for Virgo all-sky continuous wave (CW) searches. The search for this kind of signals, emitted by spinning neutron stars, covers a large portion of the source parameter space and consists of several steps organized in a hierarchical analysis pipeline. CNAF has been mainly used for the ``incoherent'' stage, based of a particular implementation of the Hough transform, which is the heaviest part of the analysis from a computational point of view. The code implementing the Hough transform has been written in such a way that the exploration of the parameter space can be split in several independent jobs, each covering a range of signal frequencies and a portion of the sky. This is an embarrassingly parallel problem, very well suited to be run in a distributed computing environment. The analysis jobs have been run using the EGI UMD grid middleware, with input and output files stored in a StoRM-based Storage Element at CNAF. Candidate post-processing, consisting of clusterisation, coincidences and ranking, and parts of the candidate follow-up analysis have been also carried on at CNAF. Typical Hough transform jobs needs about 4GB of memory (with a fraction requiring more, up to 8GB). Past year most of the resources have been used to analyze Advanced LIGO O2 data. Overall, in 2018 more than 10M CPU hours have been used at CNAF for CW searches, by running O($10^5$) jobs, with duration from a few hours to ~3 days.
\subsubsection{cWB pipeline}
Starting in 2019, the coherent WaveBurst based pipelines have been ported and adapted to run at CNAF to reproduce the cWB environment setup on the worker nodes, without the constraint to read the user home account during running. It is planned to run at CNAF all Virgo offline long duration all-sky searches on the data that will be collected during the Observational Run 3 (03) that started April 1st, 2019. cWB is a data-analysis tool to search for a broad range of gravitational-wave (GW) transients. The pipeline identifies coincident events in the GW data from earth-based interferometric detectors and reconstructs the gravitational wave signal by using a constrained maximum likelihood approach. The algorithm performs a time-frequency analysis of the data, using wavelet representation, and identifies the events by clustering time-frequency pixels with significant excess coherent power. The likelihood statistics is built as a coherent sum over the responses of different detectors and estimates the total signal to noise ratio of the GW signal in the network. The pipeline splits the total analysis time into sub-periods to be analyzed in parallel jobs, using HTCondor tools and it is expected to use a consistent amount of CPU hours during 2019.
\subsubsection{Newtonian noise pipeline}
The cancellation of gravitational noise from seismic fields will be a major challenge both from theoretical and computational point of view, since the involved simulations are very demanding. This activity requires the accurate positioning of a large number of seismometers. A cluster at CNAF was used to run position optimisations of the seismic arrays used for cancellation and to determine the cancellation performance as a function of the number of sensors and its robustness with respect to sensor-positioning accuracy.
\subsection{outlook}
The first detection of gravitational waves (GW) and the birth of multi-messenger astrophysics have opened a new field of scientific research. With the possibility to detect GW from various kind of sources we can probe new physical phenomena in regions of the Universe we couldn't explore before, with new perspectives on our knowledge about how it works.
Indeed, so far only signals from the coalescence of compact objects have been detected, while one of the most interesting and promising class of continuous GW signals, coming from asymmetrical rotating neutron stars, is still missing. Wide searches of this kind of signals require a huge amount of computational power due to the Doppler effect of the Earth motion, which disrupts the incoming signal dramatically increases the parameters space. This means that it is necessary to develop complex algorithms to reduce the computational power needed, at the price of significantly reducing the sensitivity of the search.
The development of new algorithms, which use the high efficiency and computational power of modern GPUs, showed that the new codes on a single GPU can run with a factor of ten speed-up with respect to the older ones on a ten times more expensive multi-core CPU.
For the CW case, using real data from the 9 months long run of the LIGO detectors we have estimated that on a cluster of about 200 GPUs a complete search can be done in about a couple of months, to be confronted with the several months required by the older code on a 2000 CPUs cluster.\\ A GPU cluster would be also extremely useful to test and train Machine Learning algorithms, which in the recent years were shown to be able to face very complex analyses with high efficiency and speed.\\
Advanced Virgo and Advanced LIGO are also exploring different technologies to face the new challenges of GW physics. The growing number of computing centers involved in GW research forces us to relax our idea on computing, searching a way to uniformly run different pipelines in complex and heterogeneous infrastructures. For example, the de-supporting of GridFTP pushes towards the use of Rucio, a well supported and flexible tool for data-transfer and management, while the de-supporting of the Cream-CE suggests a redesign of the job submission strategy, possibly under the control of an overall management system like DIRAC. \\ CNAF staff is intensively supporting Virgo members in all this these tests.
\end{document}