kip ivan kisel, uni-heidelberg, rt20031 22 may 2003 a scalable 1 mhz trigger farm prototype with...
DESCRIPTION
KIP Ivan Kisel, Uni-Heidelberg, RT May 2003 Trigger Concept Reduce rate from 1 MHz to 40 kHz Send data to RU (8 kB/evt 8 GB/s): VELO + TT + L0DU T1—3 (~5—10% on CPU demand) Traffic shaping --> use Scheduler ! NIC with Remote Direct Memory Access ! Prototype : 2D torus with 32 dual nodes at 1.24 MHz Trigger farm : 3D torus with up to 1200 CPUs TagNet Data PC farm (2D torus) Scheduler - RU - CN X->Y routing x yTRANSCRIPT
Ivan Kisel, Uni-Heidelberg, RT2003 122 May 2003
KIPKIP
A Scalable 1 MHz Trigger Farm Prototype A Scalable 1 MHz Trigger Farm Prototype with Event-Coherent DMA Inputwith Event-Coherent DMA Input
V. Lindenstruth, D. Atanasov, V. Lindenstruth, D. Atanasov, I. KiselI. Kisel, A. Walsch, A. Walsch( KIP, Uni-Heidelberg, Germany )( KIP, Uni-Heidelberg, Germany )
H. Muller, D. Altmann, A. Guirao, F. Vinci dos Santos H. Muller, D. Altmann, A. Guirao, F. Vinci dos Santos ( CERN, Geneva, Switzerland )( CERN, Geneva, Switzerland )
•LHCb Level-1 Trigger•Trigger Concept•Trigger Prototype•Trigger Simulation•Trigger Algorithm
Ivan Kisel, Uni-Heidelberg, RT2003 222 May 2003
KIPKIP
ee
hh
Level-1 Trigger for LHCbLevel-1 Trigger for LHCb
1. Find VELO 2D tracks and reconstruct 3D primary vertex2. Reconstruct high-impact parameter tracks in 3D3. Extrapolate to TT through small magnetic field PT4. Match tracks to L0 muon objects PT and PID5. Select B–events using impact parameter and PT information6. Use T1—3 data to improve further selection (5—10% of events)
VELOVELO TTTT (T1—3)(T1—3)
Ivan Kisel, Uni-Heidelberg, RT2003 322 May 2003
KIPKIP Trigger ConceptTrigger ConceptReduce rate from 1 MHz to 40 kHz
Send data to RU (8 kB/evt 8 GB/s): VELO + TT + L0DU T1—3 (~5—10% on CPU demand)
Traffic shaping --> use Scheduler !Traffic shaping --> use Scheduler ! NIC with Remote Direct Memory Access !NIC with Remote Direct Memory Access ! PrototypePrototype : 2D torus with 32 dual nodes at 1.24 MHz: 2D torus with 32 dual nodes at 1.24 MHz Trigger farm Trigger farm : 3D torus with up to 1200 CPUs: 3D torus with up to 1200 CPUs
TagNet
Data
PC farm (2D torus)
Scheduler
- RU- CN
X->Y routing
x
y
Ivan Kisel, Uni-Heidelberg, RT2003 422 May 2003
KIPKIP SchedulerScheduler - Basic Block Diagram - Basic Block Diagram
The supervisor of the system. Handle a coherent data transfer between RUs and CNs. Feed the TagNet with tags for synchronous data transfer in the RUs.
Scheduler
Core
TagOutputStage
TagNetFeed
TagNetFeedback
User Control
Event entries
FreeCN IDsentries
List of free computing
nodes
ControlRegister
StatusRegister
TagInputStage
Ivan Kisel, Uni-Heidelberg, RT2003 522 May 2003
KIPKIP
Tag In Tag Out
Data In
Sub
even
tB
uffe
r
DMA
PCI BusNIC Interface
NIC OutNIC In
C/M ?
CommandExecution
MessageExecution MUX Tag
Buffer
Readout Unit (RU)Readout Unit (RU)Command Tag Message Tag
Ivan Kisel, Uni-Heidelberg, RT2003 622 May 2003
KIPKIP Trigger Farm Prototype in HeidelbergTrigger Farm Prototype in Heidelberg
>1 MHz
64 CPUs
2D torus
6 Gbit/s NIC
1 year
480 MB/s p-p
450 MB/s x-y
Ivan Kisel, Uni-Heidelberg, RT2003 722 May 2003
KIPKIP
•Automatic setup of the compute farm
•Configure and control processes on every CN
GUI of PrototypeGUI of Prototype
Ivan Kisel, Uni-Heidelberg, RT2003 822 May 2003
KIPKIP
Scheduler
TagNet 3D Core 1D Cover
Data
•TagNet – schedule and send small data packets•Core network – distribute data to the target compute nodes•Cover network – increase number of compute nodes•X->Y->Z routing path
- RU- CN
3D Torus Topology 4x4x(1+2+1)3D Torus Topology 4x4x(1+2+1)
x
y
z
Ivan Kisel, Uni-Heidelberg, RT2003 922 May 2003
KIPKIP Ptolemy II Simulation of the Trigger Ptolemy II Simulation of the Trigger 3D Torus (6x6x8)3D Torus (6x6x8) 275 CNs275 CNs
Ivan Kisel, Uni-Heidelberg, RT2003 1022 May 2003
KIPKIP Simulation of the Trigger --- ResultsSimulation of the Trigger --- Results
128 B/RU 2.1 MHz measured !
1200 CPU
+5% T1--3
CompactCompact
ScalableScalable
FastFast
Fast ResponseFast ResponseVELO
ZVELO events
VELO T1-3 New VELO
Scheduler
Ivan Kisel, Uni-Heidelberg, RT2003 1122 May 2003
KIPKIPTracking Efficiency and PV ResolutionTracking Efficiency and PV Resolution
Track subsetsTrack subsets
Reference B longReference B longReference prim. longReference prim. longReference BReference primaryReference setAll setExtra setCloneGhost
97.797.799.199.196.698.797.093.681.1 4.5 6.3
95.195.197.197.193.393.992.387.570.2 4.0 9.3
2D % 3D<70> <8>
ZZcorecore 46 46 mm
X/YX/Ycorecore 17 17 mm
Ivan Kisel, Uni-Heidelberg, RT2003 1222 May 2003
KIPKIP Trigger PerformanceTrigger Performance
time (ms)
E
vent
s
17
17
ms
ms
Mean: 15 15 ss
Max: ~130 s
CPUCPU4.8 ms4.8 ms
1) Tracking efficiency 97—99% 2) PV resolution 46 m 3) Timing 4.8 ms
Expect a factor 7—8 in CPU power in 2007(PASTA report)
=> we are already within 1 ms !=> we are already within 1 ms !
•Cellular Automaton algorithm•FPGA co-processor at 50 MHz•8 processing units running in parallel
=> 15 => 15 s !s !
FPGA FPGA co-processorco-processor
E
vent
s
time (s)
Ivan Kisel, Uni-Heidelberg, RT2003 1322 May 2003
KIPKIP
Summary:Summary:
Demonstrated Architecture with 3D torus and TagNet Prototype of 64 CPUs has shown stable work at > 1 MHz The Simulation is based on the prototype measurements The Algorithm has high performance on tracks and vertices The Cost is 1300 kCHF (500 CPU) / 2300 kCHF (1000 CPU) The Team --- Heidelberg, CERN and Dubna