120 likes | 250 Views
Configuration Management for High-Performance Cluster for Real-Time Computing (ALICE High Level Trigger). Lars Christian Raae. Supervisor: Håvard Helstrup. The ALICE High Level Trigger (HLT). Trigger: A mechanism to determine if the ” record ” button should be pushed .
E N D
Configuration Management for High-PerformanceCluster for Real-Time Computing(ALICE HighLevel Trigger) Lars Christian Raae Supervisor: Håvard Helstrup
The ALICEHighLevel Trigger (HLT) • Trigger: A mechanism to determineifthe ”record” buttonshould be pushed. • HLT objective: Cutdatastream from 25GB/s to more manageable 1.2 GB/s by • finding ”interesting” events • selectingevent regions ofinterest • data compression • HLT computingcluster: On-site, COTS machines, about 1 000 CPUs
HLT ConfigurationExample T. Thingnæs. Generering av konfigurasjonsfiler for TaskManager i HLT-systemet for ALICE-eksperimentet på CERN. Master’sthesis, University of Bergen, Norway, 2007.
HLT Architecture M. Richter, Development and Integration of on-line Data Analysis for the ALICE Experiment. PhD thesis, University of Bergen, Norway, 2009. [Online] https://bora.uib.no/bitstream/1956/3555/1/Dr.thesis_Matthias%20Richter.pdf
HLT ConfigurationOptimization • Unique and complex experiment with unpredictable data stream • Initial configuration a ”qualified guess” • Configuration will need optimization for a long time • Increaseperformance by: • Minimizing node communication • Minimizing node inactivity • Eliminatingprocessingbottlenecks • Prerequisite: Test bench
Research Project • Research project: Simulationofthe HLT computingcluster • Cannotexperimentonproductioncluster • Equivalent test clustertoocostly, must usedifferent hardware configuration • Develop software solutionthat lets usmodelpreciselyenough to compareconfigurations • Openquestion: How, exactly, is thisgoing to be done?
Evaluation • What features of a real computingcluster is thesolutionable to model? • Howaccuratelycanthesolutionanswerwhich is betteroftwoclusterconfigurations, given applicable input, and howmuchtheydiffer in performance? • How portable is thesolution? Howmuch manual work is required to test a differentclustersetupthantheone given in the case?
PossibleResults • For HLT: Testbench to performexperiments and developimproved HLT configurations • For physics: Higherqualityexperiment data • For clustercomputingcommunity: Perhaps a newsimulationtool