Multi-Core R&D Project

Optimize LHC physics software applications to run on multi-core cpus

This project is part of the CERN SFT group in the PH DIvision and conducted in collaboration with the OpenLab and the LHC experiments

The project ended in 2011. A new project Concurrent Programming Models and Frameworks has been initiated that will take-over the activities carried during this R&D.

General

• Meetings
• mailing-list archive
• Action Items
• Blogs
• In case of further questions

Latest Additions

• Reducing memory footprint using jemalloc
• Reconciling OOD with DOD
• Auto-vectorize trigonometric and trascendental functions

"Recent" Summary Reports

• HEP software on Multicore: present experience and perspectives: given at the SuperB Computing R&D Workshop 2011
• collection of presentations and reports on the activities in year 2010
  • Please refer to these reports for a detailed description of the activities listed in the sections below
• Report (pdf) at the 2nd Workshop on adapting applications and computing services to multi-core and virtualization (2010)
• Report from the Workshop on "CHALLENGES FOR UNDERSTANDING THE QUANTUM UNIVERSE AND THE ROLE OF COMPUTING AT THE EXTREME SCALE"
• Report (pdf) at first Workshop on adapting applications and computing services to multi-core and virtualization (2009)
• Report (pdf) at the Kickoff Workshop on multi-core and virtualization (2008)

"Recent" Event of interest

• Workshop on Concurrency in the many-Cores Era, November 2011
• ATLAS workshop on Future computing in particle physics: June 2011
• An introduction to OpenCL: tutorial by Tim Mattson (Intel Labs)
• Workshop on HPC on Hybrid platforms at INRIA
• SuperB Computing R&D Workshop 2011

Track 1:

Objective

• Investigate current and future multi-core architectures.
• Evaluate tools to measure performance.
• Develop a measurement and analysis methodology.

Deliverables

• Assessment of industry trend in multi-core architectures.
  • see reports above
  • google for Nehalem, SandyBridge, Buldozer,, NuMa, multicore, manycore, gpgpu
• Recommendations on tools, metrics and methodology to assess the performance of LHC physics application software on such architectures
  • CMSSW Performance Monitoring on processors based on the Intel Core and Nehalem Microarchitectures by Daniele Kruse
  • see OpenLab reports and lectures at INFN school referenced below
• Instrumentation software
  • Perfmon2 by OpenLab
  • igprof by Giulio Eulisse & Lassi Tuura
  • Gaudi Profiler
  • CMSSW Profiler

Current/Recent Activities

• Investigate core i7 architecture and benchmark it with particular attention to hyperthreading (report by VI)
• See also OpenLab documents
• Establish a measurement methodology based on perfmon2
  • * CMSSW Performance Monitoring on processors based on the Intel Core and Nehalem Microarchitectures by Daniele Kruse
• Instrument LHC software (CMSSW, Gaudi, Geant4)
• Performance Analysis on Intel® Corei7 by David A Levinthal (Intel)
• ptuview by Matti Kortelainen

Previous works

• Analysing CMS software performance using IgProf, OProfile and callgrind by Lassi Tuura, Giulio Eulisse and VI
• The investigation of CMS software through an automated performance measurement and analysis system. by Danilo Piparo, Robin Moser and VI

Track 2:

Objective

• Measure and analyze performance of current LHC physics application software on multi-core architectures
• Identify bottlenecks
• Prototype solutions at the level of system and core libraries

Deliverables

• Reports on performance of current LHC physics application software
  • CMSSW Performance Monitoring on processors based on the Intel Core and Nehalem Microarchitectures by Daniele Kruse
  • OpenLab documents
• Recommendations on best practices to avoid bottlenecks and best exploit multi-core architectures
  • courses given at the INFN school on "Architectures, tools and methodologies for developing efficient large scientific computing applications" in 2009, 2010
• Eventual materialization in software library components to implement them
  • ksm.tar.bz2: ksm module for RHEL 5.2 (version march 2009)
  • modified version of the cephes library to support auto-vectorization in GCC

Current/Recent Activities

• Evaluation of tcmalloc and huge-pages report by VI
• Evaluation of KSM (report by VI, and consolidated evaluation on march 2009)
• Workshop with Intel
• Evaluation of jemalloc as a mean to reduce the memory footprint on multicore

Track 3:

Objective

• Investigate solutions to parallelize current LHC physics software at application framework level
• Identify reusable design patterns and implementation technologies to achieve parallelization
• produce prototypes

Deliverables

• Recommendations on reusable design patterns and implementation technologies to use to achieve parallelization:
  • Study on shared memory for the LHC using Boost Interprocesss library
  • G4report-11-04.txt: Progress report on multithreaded Geant4
  • see also conference reports referenced in the "activity summary" above
• Eventual materialization in software library components to implement them
  • Gaudi parallel

Current/Recent Activities

• parallelization of Gaudi using python
• PROOF-lite
• parallelization of ATLAS event processing framework
• parallelization of CMS event processing framework

Track 4:

• Investigate solutions to parallelize algorithms used in current LHC physics application software
• Identify reusable design patterns and implementation technologies to achieve effective high granularity parallelization
• produce prototypes

Deliverables

• Recommendations on reusable design patterns and implementation technologies to use to achieve effective high granularity parallelization
  • Reconciling OOD with DOD

• Eventual materialization in software library components to implement them
  • Parallel Minuit2 (google for it)

Current/Recent Activities

• Vectorization in reconstruction Algorithms (by VI and CMS team)
• port of RootFit to GPU (by OpenLab)
• Evaluation of maximum likelihood fits on GPU devices using CUDA by Felice Pantaleo. Given at "Facing the Multicore Challenge II" workshop in Karlsruhe, September 2011

-- VincenzoInnocente - 21 Oct 2007

-- MarcMagransDeAbril - 04 Jul 2008