CANFAR+Skytree: A Cloud Computing and Data Mining System for Astronomy
aa r X i v : . [ a s t r o - ph . I M ] D ec **Volume Title**ASP Conference Series, Vol. **Volume Number****Author** c (cid:13) **Copyright Year** Astronomical Society of the Pacific CANFAR + Skytree: A Cloud Computing and Data Mining Systemfor Astronomy
Nicholas M. Ball National Research Council Canada, 5071 West Saanich Road, Victoria, BCV9E 2E7
Abstract.
To-date, computing systems have allowed either sophisticated analysisof small datasets, as exemplified by most astronomy software, or simple analysis oflarge datasets, such as database queries. At the Canadian Astronomy Data Centre,we have combined our cloud computing system, the Canadian Advanced Network forAstronomical Research (CANFAR), with the world’s most advanced machine learningsoftware, Skytree, to create the world’s first cloud computing system for data mining inastronomy.CANFAR provides a generic environment for the storage and processing of largedatasets, removing the requirement for an individual or project to set up and main-tain a computing system when implementing an extensive undertaking such as a sur-vey pipeline. 500 processor cores and several hundred terabytes of persistent storageare currently available to users, and both the storage and processing infrastructure areexpandable. The storage is implemented via the International Virtual Observatory Al-liance’s VOSpace protocol, and is available as a mounted filesystem accessible bothinteractively, and to all processing jobs. The user interacts with CANFAR by utilizingvirtual machines, which appear to them as equivalent to a desktop. Each machine isreplicated as desired to perform large-scale parallel processing. Such an arrangementenables the user to immediately install and run the same astronomy code that they al-ready utilize, in the same way as on a desktop. In addition, unlike many cloud systems,batch job scheduling is handled for the user on multiple virtual machines by the Condorjob queueing system.Skytree is installed and run just as any other software on the system, and thus actsas a library of command line data mining functions that can be integrated into one’swider analysis. Thus we have created a generic environment for large-scale analysis bydata mining, in the same way that CANFAR itself has done for storage and processing.Because Skytree scales to large data in linear runtime, this allows the full sophisticationof the huge fields of data mining and machine learning to be applied to the hundreds ofmillions of objects that make up current large datasets.We demonstrate the utility of the CANFAR + Skytree system by showing scienceresults obtained, including assigning photometric redshifts to the MegaPipe reductionsof the Canada-France-Hawaii Telescope Legacy Wide and Deep surveys. This projectinvolves producing, handling, and running data mining on, a catalog of over 13 billionobject instances. This is comparable in size to those expected from next-generationsurveys, such as the Large Synoptic Survey Telescope.The CANFAR + Skytree system is open for use by any interested member of theastronomical community.
1. CANFAR
The Canadian Advanced Network for Astronomical Research (CANFAR; Gaudet et al.2011) is the cloud computing system of the Canadian Astronomy Data Centre (CADC).It is the first system designed to provide this capability to astronomers. It provides500 processor cores, up to 32G memory per processor node, and several hundredterabytes of storage, implemented via the International Virtual Observatory Alliance-compliant VOSpace system. VOSpace can be accessed both interactively and in batchas a mounted filesystem, via VOFS . The user interacts with the system via a VirtualMachine, and submits batch processing jobs via Condor. The upshot is that one canconfigure and run one’s own code as on a desktop, then replicate it across the cloud.Other science clouds exist, such as the LHC Grid , and the EU Helix Nebula project ,but none provides the facilities for astronomers of CANFAR, or the analytics capabilityof Skytree.
2. Skytree
Skytree is the world’s most advanced machine learning software. It acts as a machinelearning server to allow advanced data mining on large data (Figure 1), e.g., withinone’s data processing pipeline, or more specialized science project. Skytree’s A. Grayalso heads the FASTlab group at the Georgia Institute of Technology. The group holdsseveral records for the fastest implementation of well-known machine learning algo-rithms. Algorithms that otherwise scale as, e.g., N , for N objects, are implementedto scale linearly, without loss of accuracy. While each specific use case will remainscience-driven, the underlying tools are not dataset-specific. Thus, the installation ofSkytree on the CADC infrastructure makes possible the practical use of these algo-rithms by astronomers who are not data mining specialists. The software’s quality androbustness renders it suitable for publication-quality research.
3. Performance
Skytree claims to make fundamental machine learning algorithms scalable. We havetested these claims on the CANFAR + Skytree system, and have verified that it scales tothe largest current surveys, and beyond to upcoming datasets. To do this, we analyzelarge astronomical catalogs. An example benchmark scaling is shown in Figure 2. http://canfar.phys.uvic.ca http://canfar.astrosci.ca/wiki/index.php/VOSpace_filesystem http://wlcg.web.cern.ch ANFAR + Skytree 3
Figure 1. Skytree allows both advanced analytics, and its application to large data.This fills a vacant parameter space that is essential for future astronomical data anal-ysis.Figure 2. Typical performance benchmarks. Linear runtime scaling with datasetsize: Skytree’s allkn has computed the 5 nearest neighbors of each dataset point.A na¨ıve implementation would be quadratic, and the best fit line would have a slopeof order 2, not 1. Runtime error bars are less than the size of the points on the plot,with the exception of the 0.2 and 0.5 fractions, for which they are not yet computed.These two utilize CANFAR nodes with 256G memory.
Nicholas M.Ball
4. Science Example
We have computed photometric redshifts for Canada-France-Hawaii Telescope LegacySurvey (CFHTLS). Most studies use single values, or a Gaussian approximation. Thisis in general unsuitable, and full probability density functions (PDFs) in redshift pro-duce superior results (Ball et al. 2008). Skytree’s allkn and kde allow us to producethese PDFs for the CFHTLS nonparametrically, utilizing the full information within thetraining set. This involves handling a catalog of 13 billion objects, comparable in sizeto upcoming Large Synoptic Survey Telescope project.
5. Conclusions
CANFAR + Skytree represents world’s first cloud computing system for data miningin astronomy, and is open for use by any interested member of the astronomical com-munity. For further details on usage, see the ADASS XXII Focus Demo (Ball 2012, thisvolume), or visit the CANFAR + Skytree website at https://sites.google.com/site/nickballastronomer . Acknowledgments.
This research used the facilities of the Canadian AstronomyData Centre, operated by the National Research Council of Canada with the support ofthe Canadian Space Agency. Funding for CANFAR was provided by CANARIE viathe Network Enabled Platforms Supporting Virtual Organisations program. The authorthanks D. Schade, A. Gray and M. Hack for their contributions to this work.