high-rate level-1 trigger design proposal for the cbm experiment
DESCRIPTION
High-Rate Level-1 Trigger Design Proposal for the CBM Experiment. Ivan Kisel for Kirchhoff Institute of Physics, Uni-Heidelberg, Germany Laboratory of Information Technologies, JINR, Dubna, Russia. Level-1 Trigger: Concept Prototype Simulation Reconstruction. Objectives. - PowerPoint PPT PresentationTRANSCRIPT
Ivan Kisel JINR-GSI meeting 120-21 Nov 2003
KIPKIP
High-Rate Level-1 Trigger Design ProposalHigh-Rate Level-1 Trigger Design Proposalfor the CBM Experimentfor the CBM Experiment
Ivan KiselIvan Kiselforfor
Kirchhoff Institute of Physics, Uni-Heidelberg, GermanyLaboratory of Information Technologies, JINR, Dubna, Russia
Level-1 Trigger:•Concept•Prototype•Simulation•Reconstruction
Ivan Kisel JINR-GSI meeting 220-21 Nov 2003
KIPKIP ObjectivesObjectivesObjectivesObjectives
1. High performance farm test bed• Test most of aspects of the system using detector mock-ups.
2. MHz cluster resource management Resource management, scheduling algorithm and infrastructure.
3. MHz reliable low-cost networking framework (ATOLL, SCI, Infiniband, …) At least two NIC candidates are planned to be evaluated. At least one NIC will then be
implemented in the prototype.4. Cluster fault tolerance framework
Automatic remedy to most error conditions without human intervention or isolation and documentation for irreparable errors.
5. MHz cluster simulation• Test all possible operating aspects as well as scalability of the system.
6. Level-1 reconstruction algorithm High speed algorithm for triggering.
7. FPGA co-processor framework FPGA co-processor will implement bus snooping techniques.
We propose the development of a generic, modular, high rate, high throughput, reliable commodity compute farm infrastructure and prototype for the specific requirements of the CBM experiment. The scalability is demonstrated by appropriate simulations. The prototype farm is to be subjected to regular use by the collaboration for simulation in parallel to the real-time on-line performance testing and monitoring.
Ivan Kisel JINR-GSI meeting 320-21 Nov 2003
KIPKIP SketchSketch of of Data Flow and Data TopologyData Flow and Data TopologySketchSketch of of Data Flow and Data TopologyData Flow and Data Topology
IT
TOF
ReactionCounter
RICH
TRDTrackletsearch
Clustersearch
Clustersearch
Readout
L1 TMU
L2Algorithm
HLT / DAQ
Ringsearch
L1L1 L2L2 DAQDAQ
IT-VertexProcessor
Localprocessing
Sub-eventbuilding
Eventprocessing
Ivan Kisel JINR-GSI meeting 420-21 Nov 2003
KIPKIP 3D Topology3D Topology3D Topology3D Topology
RURU
SchSch
TRD
Input Data
x
y
z
RURU
RURU
RURU
RICH
IT
TOF
TRD
RICH
IT
TOF
PC FarmTagNet
Ivan Kisel JINR-GSI meeting 520-21 Nov 2003
KIPKIP A Compute NodeA Compute NodeA Compute NodeA Compute Node
NIC
PCI bus
CPU
In Out
ATOLLATOLL SCISCI InfinibandInfiniband…
ATOLLATOLL SCISCI InfinibandInfiniband…
FPGA
Ivan Kisel JINR-GSI meeting 620-21 Nov 2003
KIPKIP Level-1 Trigger Prototype in HeidelbergLevel-1 Trigger Prototype in HeidelbergLevel-1 Trigger Prototype in HeidelbergLevel-1 Trigger Prototype in Heidelberg
>1 MHz>1 MHz
32 dual CN32 dual CN
2D torus2D torus
6 Gb/s 2D SCI6 Gb/s 2D SCI
480 MB/s p-p480 MB/s p-p
450 MB/s x-y450 MB/s x-y
1 Gb/s Ethernet1 Gb/s Ethernet
Ivan Kisel JINR-GSI meeting 720-21 Nov 2003
KIPKIP
•Automatic setup of the compute farm
•Configure and control processes on every CN
GUI of PrototypeGUI of PrototypeGUI of PrototypeGUI of Prototype
Ivan Kisel JINR-GSI meeting 820-21 Nov 2003
KIPKIP Hardware Initiated DMA TransferHardware Initiated DMA TransferHardware Initiated DMA TransferHardware Initiated DMA Transfer
Ivan Kisel JINR-GSI meeting 920-21 Nov 2003
KIPKIPPtolemy II Simulation of the TriggerPtolemy II Simulation of the TriggerPtolemy II Simulation of the TriggerPtolemy II Simulation of the Trigger
Ivan Kisel JINR-GSI meeting 1020-21 Nov 2003
KIPKIP TRACK RECONSTRUCTIONbased on the Cellular Automaton Method
XZ (bending) / YZ (non-bending)
TRACK RECONSTRUCTIONbased on the Cellular Automaton Method
XZ (bending) / YZ (non-bending)
Ivan Kisel JINR-GSI meeting 1120-21 Nov 2003
KIPKIPTRACKING EFFICIENCYTRACKING EFFICIENCY
RECO STATISTICS 100 events Refprim efficiency : 98.36 | 46562 Refset efficiency : 94.85 | 49250 Allset efficiency : 90.09 | 64860 Extra efficiency : 77.79 | 15610 Clone probability : 0.11 | 74 Ghost probability : 5.18 | 3358 Reco MC tracks/event : 648 Timing/event : 175 ms
RECO STATISTICS 100 events Refprim efficiency : 98.36 | 46562 Refset efficiency : 94.85 | 49250 Allset efficiency : 90.09 | 64860 Extra efficiency : 77.79 | 15610 Clone probability : 0.11 | 74 Ghost probability : 5.18 | 3358 Reco MC tracks/event : 648 Timing/event : 175 ms
ALL MC TRACKSALL MC TRACKSRECONSTRUCTABLE TRACKS
Number of hits >= 3
REFERENCE TRACKS
Momentum > 1 GeV
TIMING (ms)
Fetch ROOT MC data 63.3
Copy to local arrays and sort 12.4
Create and link segments 115.7115.7
Create track candidates 53.553.5
Select tracks 2.62.6
TIMING (ms)
Fetch ROOT MC data 63.3
Copy to local arrays and sort 12.4
Create and link segments 115.7115.7
Create track candidates 53.553.5
Select tracks 2.62.6
FPGACo-processor
98%
CPU2%
CA – INTRINSICALLY LOCAL AND PARALLEL
CA – INTRINSICALLY LOCAL AND PARALLEL
Ivan Kisel JINR-GSI meeting 1220-21 Nov 2003
KIPKIP Plans:Plans:Plans:Plans:
Kirchhoff Institute of Physics, Uni-Heidelberg, GermanyLaboratory of Information Technologies, JINR, Dubna, Russia
Develop Architecture with 3D topology and TagNet. Heidelberg (3)
Develop Scheduler. Heidelberg (1)
Investigate the Prototype of 32 dual CNs at > 1 MHz. Heidelberg (2)
Advance the Simulation based on the prototype measurements. Dubna (2)
Investigate different Network Interface Cards applicability. Heidelberg (3) + Dubna (2)
Develop the Reconstruction algorithm. Heidelberg (1) + Dubna (4)