1 / 18

Status & development of the HYDRA framework

Status & development of the HYDRA framework. Jochen Markert. Status after p+Nb beam. Future experiments: Heavy ion collisions with high track multiplicity Upgrade of HADES: Replace TOFino by RPC detector New DAQ New hld Event format ~ 50 times higher data volume. Consequences.

skeegan
Download Presentation

Status & development of the HYDRA framework

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Status & development of the HYDRA framework Jochen Markert HADES collaboration meeting 2010, GSI

  2. Status after p+Nb beam • Future experiments: • Heavy ion collisions with high track multiplicity • Upgrade of HADES: • Replace TOFino by RPC detector • New DAQ • New hld Event format • ~ 50 times higher data volume HADES collaboration meeting 2010, GSI

  3. Consequences • Replacement of TOFino by RPC: • Unpacker + Calibrater + Hit -/Clusterfinder + Digitizer of RPC to added • Since RPC does not simply replace TOFino a new integration scheme for tracking is needed • Change of hld format: • New Unpackers for all detectors needed • Track multiplicity in central heavy ion collisions: • increase of combinatorics, computation time, tracking ghosts • Decrease of tracking efficiency and quality • Increased data volume: • New strategy for DST production and analysis needed HADES collaboration meeting 2010, GSI

  4. Upgrade strategy I (source code management) • Lots of code changes and clean up of 10 years of history to be done • CVS repository is closed • move to SVN repository for better management • Permissions for developers managed by database • Code can be checked out / commited to the web server • hydra (https://subversion.gsi.de/hades/hydra) • containing all history up hydra v8_21. • will not be touched besides important bug fixes. • should be used for analysis of old data (up to sep08). • hydraTrans (https://subversion.gsi.de/hades/hydraTrans). • For the transition period to the final hydra frame work. • development version. All changes of the code will be committed here. • not backward compatible to read old DST ROOT files • Should be used for simulations for the heavy ion runs • hydraII (https://subversion.gsi.de/hades/hydraII). • final repository used for the analysis of the upcoming experiments 2010 and later. HADES collaboration meeting 2010, GSI

  5. new Tasks for MetaMatch, Spline and RK (later labeled "new"). • new pid/pair to have different container classes and therefore a new Filler + Sorter/Cleaner. • Old analysis with ShowerTofino will not be supported by the new data structure. HADES collaboration meeting 2010, GSI

  6. Task assignment • Vladimir, Olga: Tracking • Anar: adaption of Spline/RK • Ilse : ORACLE/RuntimeDB, Unpacker integration, START detector • Jochen: common framework, replacement of old PID/PAIR • Gosia, Tatyana: Analyis development • Detector libs + Unpacker by detector groups HADES collaboration meeting 2010, GSI

  7. Time line A bit delayed…but not too much HADES collaboration meeting 2010, GSI

  8. Changes in hydraTrans • done • Remove lib: • alignment lib • mdctrackS lib • online lib • phyana lib (HPhysicsConstants moved to base/util) • showerTofino lib (remaining classes now in shower lib included) • remove GO4 , rename hadesgo4 to online lib (see Julian Thomas talk on Friday) • add new lib • startNew dir for development of new start detector (not included into Make) • oraNew (as above) • Particle (replace old pid/pairs lib) • new tracking code (not fully finished) • Integration of RPC (code for parameters to be done) • remove cross dependencies from libs • To be done • remove kickplane lib • Adopt dst lib • remove obsolete classes from all libs (partly done) • add pair framework to lib particle HADES collaboration meeting 2010, GSI

  9. Where are we now? • Full simulation chain: • RPC detector integrated • Full DST chain: • RPC integrated • New MDC tracking code • Efficiency + purity results looking promising (see Vladimir’s talk and proposal) • Tracking speed strongly improved (2 events/s fully central Au+Au collisions) • New DST output format (see next transparency) • Preliminary Lepton analysis in Au+Au available (see Tatyana’s talk on Friday) • Missing: • Iterative track finding for off-vertex tracks • Global track fitting (Kalman filter etc.) • Speed optimization HADES collaboration meeting 2010, GSI

  10. DST format considerations • Old HPidTrackCand objects are very large  Disk space • Many not needed data members … • Old theorem: “collect everything and decide at the end” not feasible anymore  too many combinations HADES collaboration meeting 2010, GSI

  11. What has to be done … • Strategy (in Tracking, see. Vladimir‘s talk): • Remove as many ghosts from inner part of the MDCs before doing combinations with outer detectors • Use TOF hits in candidate search to reduce the number of combinations • Remove “non physical” candidates before going to fit procedures • Change output format: • Keep minimum number variables for most tasks • Allow extensions to the output objects for special purposes • Reduce variable types to appropriate value range • Use compression features of ROOT • Do not write RuntimeDatabase to output file • Do not write HGeantKine category to DST output in Simulation • Enable read of GEANT info with second ROOT source if needed HADES collaboration meeting 2010, GSI

  12. Comparison of object sizes * ROOT output is compressed in addition (factor typical 0.6) HADES collaboration meeting 2010, GSI

  13. Further reduction of data volume • Float_t variables need 4 bytes, but often the precision is not needed in the DST output • Use Float16_t (truncated) • In memory 4 bytes, normal Float_t precision • Truncated when written to disk • 30 % reduced data size on disk (> 50 TB disk space) • Do not store RuntimeDB in each ROOT file • ~ 2 TB disk space HADES collaboration meeting 2010, GSI

  14. HParticleCand HParticleCand HParticleCand HParticleDebug HParticleDebug HParticleDebug HParticleCal HParticleCal HParticleCal Flexible extension of DST output … 1st object 2nd object 3rd object • For special tasks additional objects can be stored in the same order in the DST output • Objects can be easily accessed by index • Even T->Draw(….) commands will work since objects are stored exactly parallel • DST output can be setup for example like: • Do 10% of files with Calibration info for monitoring etc HADES collaboration meeting 2010, GSI

  15. Compute requirements estimation • beam intensity of 1.5 * 107 ions per 5-sec spill. • 1% interaction target •  reaction rate of 3 * 104/sec. • LVL1 trigger set to select 25% of all reactions, • 10 kHz accepted LVL1 • 4.32 * 108 semi-central LVL1 events per day on disc (assuming 50% duty factor and dead time). • maximal possible trigger rate of 20 kHz will be filled with (scaled-down) peripheral reactions HADES collaboration meeting 2010, GSI

  16. 30 days beam • Au+Au • Hld event size estimated from simulations 25kB • DST event size 9kB • Read speed 20MB/s HADES collaboration meeting 2010, GSI

  17. CPU power • Real time: 3325 CPU cores needed •  unrealistic • Compromise: • 1500 CPU cores 3 month per year for DST • 1000 CPU cores full year for analysis • Situation now: • 400 CPUs + X  600 parallel jobs • Where to find additional 1100 CPU cores ? • What about disk space (> 400 TB missing)? HADES collaboration meeting 2010, GSI

  18. Summary • Software already now in good shape • Still features missing … • GOOD MESSAGE: Hades can handle heavy ion collisions • What about the IT environment ? • Reliability of Lustre / batch farm • Missing resource: CPU + disk space • Man power …. HADES collaboration meeting 2010, GSI

More Related