220 likes | 331 Vues
HF Luminosity and Jet Triggering. Drew Baden, Tullio Grassi, Jeremy Mans University of Maryland Chris Tully Princeton University Bob Hirosky University of Virginia. HF. HF Basics. HF covers ~3< h <5 Steel absorber and rad hard quartz fibers
E N D
HF Luminosity and Jet Triggering Drew Baden, Tullio Grassi, Jeremy Mans University of Maryland Chris Tully Princeton University Bob Hirosky University of Virginia
HF HF Basics • HF covers ~3<h<5 • Steel absorber and rad hard quartz fibers • Cerenkov light collected via phototubes, uniform HCAL readout • HF+ and HF- • 36f and 12h = 432 towers per side • Df=10° and Dh=0.166 • Each tower has a long and a short fiber running along z • Short is in the back ~ “ETHAD” • Long is front to back ~ “ETEM+HAD” • Makes 2x432=864 towers per side • Level 1 Trigger and HF • Combines 2f x 3h = 6 towers = 1 trigger primitive (TPG) • Df=20° and Dh=0.5 • Only used for MET and JETS • Not for electrons and photons
Jets in Level 1 • TPGs • HB/HE: 0.087x0.087 DhxDf 0.5 x 0.35 in HF • CMS Level 1 Jets • Calorimeter organized into “Regions” • 4x4 TPGs per region in HB and HE • 1 TPG per region in HF • Jet finding via “sliding window” • Sliding window using 3x3 regions: 12x12 TPGs • Jet candidates are sent from RCT to GCT, which sorts and sends candidates to GT by category: • 12 jets, nominally 4 highest each of central, forward, and tau jet candidates Calorimeter “Region” Barrel/Endcap
Min Bias • Min bias distributions @ 1034 • <n>~20 (assumes ~80mb inelastic) • <dn±/dh> ~ <dn0/dh> ~ 8 @ 14TeV • <ET> ~ few GeV (and falls exponentially) • 20<n> x 8ch x 10dh x 2GeV = 3.2 TeV/interaction (Had) • HF...40% of CMS in Dh coverage • 640 GeV, 72 TPGs/side, ~10 GeV/TPG (@<n>=20) • Current TPG = 0.5 x 0.348 hxf • s • New Level 1 triggering will need to… • Sharpen efficiency • Move HLT-like algorithms and resolution as close to L1 as possible Atlas CERN/LHC 96-40
SLHC Background Reduction • Beat down the background L1A rate • LHC Design luminosity of 1034 has large backgrounds: • Depending on the scheme for high luminosity • <n> ~ current for 25ns SLHC, rates scale with Luminosity • for continuous beam….smaller <n> but large pileup...not sure • 100kHz L1A rate is ingrained, will most likely hold • Size of derandomizing buffers, etc. • Bandwidth to HLT • Number of HLT processors.... • CMS Calorimeter trigger based on TPGs • In HF… • 1 TPG = 6 towers (3h x 2f) • Lack of granularity might make it useless for Level 1 jet triggers with large number of multiple interactions • In HB and HE… • Jet-finding in Level 1 very inflexible, probably cannot be changed • Ditto for isolated e and g triggers • We propose to start R&D on improvements
SLHC Signal Enhancement • Add functionality • Higgs id without a tag is very hard • Gluon fusion backgrounds are too high, esp at 1035 • W Boson Fusion (WBF) dominant experimentally accessible rate • Forward jets + central Higgs decay • Tag jets are in HF+HE so HE will need to be included • Current trigger at high luminosity will be difficult • Depends on scheme for increasing luminosity of course… h (tagged “forward” jets) C. Tully & H. Pi JetMetPRS Aug 2004
Simulation Results 1034/cm2/s • Chris Tully led effort to study current trigger @ 1035/cm2/s • Chose 25ns RF structure for lack of any better guidance • Compared WBF qqH to QCD background • Study “feature bit” for HF TPG • Current version: bit = 1 if any of the 6 towers in the TPG > % of total TPG ET • Used a 2x2 instead of 3x2 for expediency • Not much discrimination at LHC or SLHC luminosities • At 1035 can see the effect of minbias/underlying event adding significant energy to jet • “feature bit” becomes “featureless”… 1035/cm2/s
A first look… • Current scheme • Jet candidates using 3x3 CR sum, DhxDf=1.5 x 1.0 • Slides window by 1 CR, DhxDf=0.5 x 0.35 • New scheme • Construct jet candidates from 4x4 tower sums, DhxDf=0.67 x 0.7 • Slide window by 1 tower, DhxDf=0.17 x 0.17 • Feature bit on if number of cells needed to sum 60% of ET (n60) in 4x4 cluster < cut • Use n60 < 7 to set feature bit • Prelim studies show QCD jets are narrow & well contained • Jets from .5 cone all have 2nd moment < 0.3 in R • Require jet candidate threshold && feature bit = 1 • Can also require perimeter “quiet” (not studied yet…)
n60 • LHC design luminosity 1034/cm2/s and SLHC 1034/cm2/s w ith 25ns RF structure… • High efficiency and background rejection for both LHC and SLHC • e > 80% and background rejection = 3 @ 1035/cm2/s 1035/cm2/s 1034/cm2/s
WBF Signal vs background • Signal: WBF qqH, MH=130 GeV • Background: QCD from uniform sampling 30 – 300 GeV • Central trigger H WW ℓnℓn (ℓ=t,e) • Add forward jet trigger using current and new algorithm • Compare 1034 and 1035 • Current algorithm jet candidtes from DhxDf=1.5 x 1.0 sliding window and “feature bit” • New algorithm jet candidates from DhxDf=0.67 x 0.7 sliding window and n60 cut • Comparison: • Signal: • Efficiency goes up in both cases due to volunteer t’s in central – this needs more study • Background: • Current algorithm: goes from 49% to 100% • New algorithm: stable at 30%
Hardware Design • We propose NOT changing current HTR • 48 QIE channels input • 6 SLB sites • Each SLB transmits 2 TPGs/twisted pair over 4 pair • For HF, 1TPG = 6 QIE channels • 48/6 = 8 TPG output • HF HTRs will only populate 1 SLB site • We will have 5 free SLB sites to use • Proposal: • Luminosity: dedicate 1 SLB site • Jet trigger: dedicate 4 SLB sites
SLB connectors Virtex2PRO FPGA Bottom view SMA I/O RCT receiver connectors RJ45 I/O Top View Luminosity • 1st prototype of combined Luminosity/SLB/RCTtest board made • 1 SLB site footprint on bottom • SLB connectivity: • 36bits from each HTR Xilinx + TTC+Localbus • UW RCT receiver footprint on top • Xilinx Virtex2PRO FPGA • Rocket I/O, embedded PPC, block ram... • Uses: • Level 1/TPG commissioning • Can host RCT Vitesse quad receiver • Will buffer TPGs and transmit back into HTR FPGA for testing • Simulating Level 1 • Has RJ45 and SMA output which can drive Jeremy’s HCAL trigger board • Luminosity • 72 bits from HTR xilinx can accommodate full 48-channel single tower threshold • Output: use RJ45 and drive gigabit ethernet • Under consideration... • Plan to meet w/Marlow&Tully soon...
HF Trigger • Topology: • Each HF HTR receives 12h x 4f towers • Need 9 HTRs per side for the long fibers • Baseline Jet algorithm • This suggests a 4x4 sliding window to contain the jet • For isolation another 2 on each side in f (and maybe also in h) • 6x6 for area for each jet candidate • What are the I/O considerations here? • Need to be able to get the data from the HF/HTRs into some kind of cluster finder
HTR/Cluster Finder I/O • It’s all in the edges... • Figure shows 3 HTR worth of data • Find 4x4 jets sliding 2 in f over each “edge” • There are 5 possible groupings of 4 towers within 2 of the edge • Heavy lines show the possible set of towers in f • For a 6x6 design, will need 3 HTRs worth of data for each cluster finder • Figure shows 2 cases: central HTR data and left+right for isolation, and a jet that straddles the border • Each HTR will have to send its 12x4 towers to 2 separate cluster finders • I/O calculations involve 48 towers • Cable plant involves 48x2 towers
12.5ns Bandwidth and Cabling • HTR Xilinx →4-slot Trigger card (HFT, for “HF Trigger”) • Needs all 48 channels, but only enough lines to send data from 32 channels @ 40MHz • Need to run the transmission at 80MHz. • 80MHz tests show that with proper termination and clock phasing, should work ok • HFT → Cluster finder • Try to send data over cat6/cat7 quad twisted pair using 8B/10B • We already know how to do 8B/10B since we are doing this now • Minimizes costs and engineering (no optical) • Will need to make the copper cables as short as possible • Need to already consider rack topology
Rack Topology • Chris Tully worked this out with Rohlf and Ianos • New “Luminosity” VME crate in the center minimizes distance from the 3 HF crates and doesn’t break Wesley’s topology for Level 1 • Goal: keep cable lengths < 5m
Data Data Data Tx Rx TTC TTC RX Clk QPLL Fifo Fifo Fifo Data Data Data Tx Rx Compensation Logic TTC TTC RX Clk Crystal QPLL QPLL QPLL I/O Scheme Using Copper • Current scheme: • Use TTC clock into the QPLL to drive our links • Asynchronous fifo at receiving end for phase synchronization • Alternate scheme: • Use crystal oscillators to drive links, and asynchronous FIFOs on TX • Latency cost in frame clock ticks (80 or 120MHz frame clocks) • Small number of frame clock ticks on TX end • Logic on receiver end – links will be not be running at LHC frequency • Would necessarily increase latency by some small number of frame clocks • Note that current RCT+GCT latency is 40 clock ticks, so we would have that amount of time to “play with” • Current HTR firmware uses ~12 clock ticks now • We think we will have plenty of time to produce jets for the global trigger
I/O Considerations Drive Everything • Current system SLB transmits using 1.2Gbps links • Fully populated HTR has links/SLB, each @ 1.2Gbps • Each SLB drives 1 quad twisted pair cable running over 10m • We plan to use 4 SLB sites • A single card would be 15.5cm high, capable of housing an 8way RJ45 connector ala the Princeton Fanout card (which has 2 of these) • So we need to run our links on 4 quad twisted pair cat6/7 cables, sending the data to 2 separate receivers • So we need to squeeze out all 48 channels onto 4 quad twisted pair (16 pair) • Need to send 3 channels/twisted pair every 25ns, or 3 bytes @ 40MHz • If we add hamming codes, that would make 4 byes @ 40MHz = 160MB/s • Using the same 8B/10B encoding as now, that makes 1.6Gbps • This is the same gigabit ethernet that we use now • Note: Xilinx FPGAs have built-in I/O exceeds 3Gbps • XC2VP30 has 8 built-in transceivers and costs ~$500 today, would need 2 per HFT • R&D is needed here. Note: da Silva uses these for the ECAL DCC….
HF Jet Board (HFJ) • This will be a 9U VME board in the “Luminosity” crate • Straightforward design • Links on RJ45 connectors, deserialized, fanout to FPGA • 9 HTR/side so try to service 3 HTR/HFJ • 8 RJ45/HLT means 24 RJ45/HFJ – HFJ would probably have to be 2 VME widths • Need 3 HFJ/side, or 6 total, or 12 VME slots • Pin gymnastics • Each of 24 RJ45 will have 4 twisted pair, or 96 pair total per HFJ, or 96 deserializers! • HTRs have 16….maybe this is doable… • Might be better off using something like the Vitesse quad deserializers • 96 deserializers, 20 pins each 1920 total parallel digital pins • Use 3 FPGA/HFJ, each HFJ services 32 inputs and needs 640 pins for I/O • Xilinx XC2V4000 flip-chip has 912 user I/O pins, there are others… • Clustering algorithms… • Needs a lot of study – simulation and hardware implementation
Optical I/O Scheme • Use optical to run faster by x4 and save connectors • 10Gigabit ethernet is becoming more and more common • Supported by the FPGA internal transceivers • Will definitely need to use crystals for transmitting • Single mode fibers are becoming more and more easy to use • Would allow single-width VME boards for the HFJ • Fewer connectors (24/HFJ down to 6 optical) • Board layout, mechanics of connectors, will be critical • R&D would have to start soon for this path
R&D Program • Scrub design so that we can piggyback on existing system and build cards that can be implemented by LHC startup • Keep entire current VME architecture, but add new capability • Run parasitically, collect data, study, iterate • R&D list for Luminosity project • Settle on Luminosity requirements, finish prototype single-SLB width board • R&D list for Trigger project • Simulation needed to settle on algorithm approach • HLT card: • Learn how to use new FPGA’s with embedded processing, DSP, built-in deserializers…. • Verify HTR to HLT @ 80MHz • HLT to HJF transmission • Study transmitting signals over <5m cat7 copper cables @ 1.6Gbps • Use crystals to drive transmitter as an alternate scheme • R&D quad serializers on the HLT and quad deserializers on HFJ • 10G optical transmission using crystals and single-mode fibers • HFJ • Study algorithms for clustering – lots of simulation needed here • R&D on how to include HE to get the meat of the WBF signal and extend jet trigger to full CMS