1. CBM Experiment Computing Topics Volker Friese GSI Darmstadt HIC4FAIR Physics Day FIAS, Frankfurt, 11 November 2014

2. Outline o A brief introduction to CBM o for more details, in particular on reconstruction issues, see the presentation at the H4F Physics Day 27 November 2013 (if you can find it on the web) o Current Topics in CBM Computing o from time slices to events o online and offline o software triggers o offline computing H4F Physics Day, Frankfurt, 11 November 2014V. Friese 2

3. CBM in a Nutshell Compressed Baryonic Matter: a heavy-ion experiment at the future facility FAIR in Darmstadt Investigation of strongly interacting matter at extreme net-baryon densities Fixed-target operation on extracted beams, 2 – 45 GeV/nucleon Large-acceptance spectrometer (dipole + Si main tracker) Hadron, lepton and photon ID: RICH, TRD, TOF, ECAL Complex event topology: up to 600 charged tracks per collisions in the acceptance Observables: yields, spectra, flow, correlations, fluctuations of bulk hadrons, multi-strange hyperons, open charm and charmonium; low-mass di-leptons Operation from 2021 on H4F Physics Day, Frankfurt, 11 November 2014V. Friese 3

4. CBM: Experimental Setup H4F Physics Day, Frankfurt, 11 November 2014V. Friese 4 HADES CBM@SIS-100 Magnet, STS + MVD RICH MUCH TRD TOF ECAL (parking) PSD

5. Computing Tasks H4F Physics Day, Frankfurt, 11 November 2014V. Friese 5 Reconstruction From raw data to tracks and particles Reconstruction From raw data to tracks and particles Simulations Transport / Detector Geometry Detector Response Digitisation Simulations Transport / Detector Geometry Detector Response Digitisation Analysis From tracks to physics observables Analysis From tracks to physics observables Online Systems Data transport Event building Data distribution and management Online Systems Data transport Event building Data distribution and management Services Experiment Control System Databases Services Experiment Control System Databases

6. CBM Readout Concept H4F Physics Day, Frankfurt, 11 November 2014V. Friese 6 Finite-size FEE buffer: latency limited throughput limited courtesy W. Müller

7. Consequences H4F Physics Day, Frankfurt, 11 November 2014V. Friese 7 The raw data has to be reduced online to a recordable rate – e.g., at 10 MHz, the raw data flux is about 1 TB/s; – selection of data based on rare observables involves complex signatures; will be performed in software only; – we need (at least partial) reconstruction in real time; – all online algorithms must be highly optimised to speed; – this requires a high level of parallelisation on modern computer architectures. The physics reach of CBM is determined by its online computing capabilities.

8. Topic: From Timeslice to Event H4F Physics Day, Frankfurt, 11 November 2014V. Friese 8 As a consequence of free-streaming data acquisition with self- triggered front-end electronics, the basic readout structure is a timeslice = container of raw data of a given time interval. – physical events (collisions) are not defined on the raw data level Timeslice: all detector data ~100 μs created in software (FlesNet) delivered to one computing node Timeslice: all detector data ~100 μs created in software (FlesNet) delivered to one computing node microslice: data from one input link ~1 μs created on FPGA (DPB) microslice: data from one input link ~1 μs created on FPGA (DPB)

9. R/O Libraries H4F Physics Day, Frankfurt, 11 November 2014V. Friese 9 A two-level approach in software: FlesNet + CbmRoot courtesy D. Hutter Under development Hardware R/O chain not yet complete Temporary solutions for current in-beam detector tests

10. Timeslice Reconstruction H4F Physics Day, Frankfurt, 11 November 2014V. Friese 10 Reconstruction starts from timeslice, but analysis deals with events: where is the transition? raw data cluster hits tracks vertices raw data cluster hits tracks vertices Timeslice Event Optimal solution may depend on running conditions (interaction rate). ? ?

11. “Event Building” on Raw Data H4F Physics Day, Frankfurt, 11 November 2014V. Friese 11 Advantage: Simple, fast, available All present event-based reconstruction algorithms can be used Needed: Tagging (or rejection) of possible pile-up events Will work only up to a certain interaction rate (~1 MHz) cand. 1cand. 2 cand. 3

12. “Event Building” on Tracks H4F Physics Day, Frankfurt, 11 November 2014V. Friese 12 Advantage: event definition easier on track level than on hit level Missing: time-based hit preparation (cluster finder, hit finder) Missing: interfaces to common framework 4D CA Track Finder: combines hits only if close in time; reconstructs track start time courtesy V. Akishina

13. Topic: Online and Offline H4F Physics Day, Frankfurt, 11 November 2014V. Friese 13 Paradigm: a common environment for online and offline software – each algorithm must (formally) run both online and offline. Easy to state, hard to realise There will be reconstruction and analysis tasks which are specific for offline (not needed for online data selection). What is the proper environment / framework? – “traditional” CbmRoot has limitations: not ideally suited for time-based simulation and reconstruction no concurrency model performance

14. ALFA and FairMQ H4F Physics Day, Frankfurt, 11 November 2014V. Friese 14 Very similar problems with ALICE R3 upgrade (continuous TPC readout, online reconstruction needed) ALFA: Alice-Fair-Framework (for online and offline) Major ingredient: FairMQ (asynchronous messaging toolkit) – process distribution, inter-process communication (socket API) – allows flexible configuration (devices, publish-subscribe, push-pull, request-reply) courtesy M. Al-Turany

15. A First Example for CBM H4F Physics Day, Frankfurt, 11 November 2014V. Friese 15 courtesy N. Winckler

16. A First Example for CBM H4F Physics Day, Frankfurt, 11 November 2014V. Friese 16 courtesy N. Winckler

17. FlesNet and FairMQ H4F Physics Day, Frankfurt, 11 November 2014V. Friese 17 FairMQ provides a development environment for parallel reconstruction software. It can be used on a single computing node (starting from FlesTimeSlice delivered by FlesNet). But also provides a model for software DAQ – concurrent to FlesNet. Decision must be based on performance and stability.

18. Topic: Software Triggers H4F Physics Day, Frankfurt, 11 November 2014V. Friese 18 Current online approach: reconstruct full event, evaluate trigger signature, discard or select data But: trigger signatures are manifold: J/psi -> muons: first-level decision in last layer of MUCH only Anti-hyperons, multi-strange objects: decision requires STS + TOF (anti- proton) only (no decay topology) Hyperons, hyper-nuclei, exotics: decay topology in STS Open charm: decay topology in MVD Many observables without trigger signature (di-electrons)

19. Software Trigger Architecture H4F Physics Day, Frankfurt, 11 November 2014V. Friese 19 In order to make optimal use of the available online computing resources, a one-step approach is probably not optimal. The trigger software must be flexible to cope with different trigger signatures: multi-step data reduction multiple triggers in parallel, including a random minimum-bias trigger flexible, easy-to-change configuration for different setups / running conditions Always reconstructing STS in the first step may not be the wisest choice. A monolithic online code is not a good development model.

20. Topic: Offline Computing H4F Physics Day, Frankfurt, 11 November 2014V. Friese 20 CBM will produce several PB per run campaign (after online data selection). Efficient analysis of such large data sets is not trivial. Distributed computing will be necessary, but will look different from the LHC grid approach. A FAIR offline computing model is under development in cooperation with GSI/FAIR-IT and the FAIR experiments. see presentation by T. Kolleger

21. Summary / Outlook H4F Physics Day, Frankfurt, 11 November 2014V. Friese 21 as usual: a long road still to go – but we are under way!