Efficient Implementation of High-Energy Physics Processing for the RCT SLHC Upgrade Using Modified Jet Reconstruction and Active Size Partitioning Anthony Gregerson Electrical & Computer Engineering Department University of Wisconsin-Madison Madison, WI [email protected] Abstract—The upcoming upgrades to the Large Hadron Collider – the world’s largest particle accelerator – and its associated experiments will require a significant redesign of its triggering electronics, including the implementation of new algorithms in the Regional Calorimeter Trigger (RCT). In this project, I intend to explore two aspects of the implementation of the RCT upgrade: (1) The algorithmic tradeoffs available in the RCT’s Jet Reconstruction function and their effect on the hardware implementation cost and (2) the impact of different using existing chip partitioning schemes compared to an algorithm-aware partitioning scheme developed for the RCT. I. INTRODUCTION Modern high-energy physics (HEP), the study of particle interactions at high energy levels, relies heavily on high-throughput, real-time DSP systems that are used to process data from experiments conducted using particle accelerators. A critical function of such systems is to perform triggering, a process that is used to reduce the large amounts of data produced by HEP experiments to a manageable level. Trigger systems achieve this goal by rapidly processing sensor data in real time and selecting which data to save and which to discard. This process typical involves analysis of raw sensor data and the reconstruction, identification, and sorting of particles and jets1. The triggering systems at the Large Hadron Collider (LHC) have stringent requirements in terms of both throughput (100s of GB/s) and real-time processing deadlines (sub-µs). To achieve these goals, the existing trigger systems are large-scale, consisting of tens of thousands of ASICs and FPGAs [1]. The Regional Calorimeter Trigger (RCT) for the Compact Muon Solenoid (CMS) experiment alone requires approximately 300 PCBs containing several thousand ASICs. The RCT is due to be redesigned and upgraded to meet the needs of the Super Large Hadron Collider Phase I upgrade project [2] [3]. These upgrades may require new algorithms with higher resolution jet reconstruction and particle isolation [4] [5]. These algorithms are described in more detail in Section II. 1 Large sprays of particles with a common source. Analysis of the implementation and partitioning of the new RCT algorithms is important for several reasons. First, the existing system is already very large scale, complex, and expensive, therefore reducing the hardware implementation cost through algorithmic implementation choices and design partitioning is an important priority. Second, the current implementation is partitioned in a way that requires a significant amount of inter-chip communication. This high level of inter-chip communication proved to be a significant challenge in the commissioning of the hardware [6]. Third, the nature of hardware triggering algorithms lends itself to algorithmic modification cost-benefit analysis. Triggers are not tied to exact, fixed algorithms; rather they must achieve satisfactory trigger efficiency – a measure related to the percentage of physics objects that are correctly identified and saved by the trigger algorithm. By providing the physicists with several possible implementation choices and their associated hardware costs, they are better able to judge the tradeoffs of targeting different efficiency levels. Additionally, as part of the upgrade process, the RCT implementation platform will shift from the custom ASICs used in the original system to a new FPGA-based platform. Due to the different nature of ASICs and FPGAs, some implementation approaches used in the previous system may no longer be efficient in the new system [7]. II. REGIONAL CALORIMETER TRIGGER ALGORITHM FOR THE SLHC PHASE I UPGRADE The RCT algorithm produces two groups of outputs: • An array of 4096 Particle objects. Each object is composed of a particle type (2 bits), particle energy (11 bits), center of energy (4 bits), and valid bit (1 bit). • An array of 4096 Jet objects. Each object is composed of a jet energy (15 bits) and valid bit (1 bit). Additionally, both jets and particles have an additional 12 bits of position information however this information is 'encoded' into the hardware – i.e. associated with specific wire locations – throughout most stages of the algorithm. The RCT algorithm has three groups of outputs:• A 2-dimensional array of 4096 electromagnetic energy readings, ECAL[1..n][1..m]. Each element is 8 bits. • A 2-dimensional array of 4096 hadronic energy readings, HCAL[1..n][1..m]. Each element is 8 bits. • A 2-dimensional array of 4096 finegrain bits, fg[1..n][1..m]. Each element is 1 bit. The input arrays have identical dimensions and are spatially related. Together the overlapping elements from the three arrays can be combined to form an array of 4096 elements, towers[1..n][1..m], where each tower is a 17-bit element composed of one ECAL one HCAL, and one fg. The RCT algorithm contains several computation stages. These stages are described in pseudocode and mathematical notation in the sections that follow. A. Tower Filtering Tower filtering is applied to all input elements. The operation zero's out each tower if both its ECAL and HCAL elements do not meet a programmable threshold value. for(i = 1 to n) do for(j = 1 to m) do if((ECAL[i][j] > ECAL_threshold) && (HCAL[i][j] > HCAL_threshold)) { ECAL_filtered[i][j] = ECAL[i][j]; HCAL_filtered[i][j] = HCAL[i][j]; fg_filtered[i][j] = fg[i][j]; } else { ECAL_filtered[i][j] = 0; HCAL_filtered[i][j] = 0; fg_filtered[i][j] = 0; } B. Channel Energy Sums The sums of ECAL and HCAL for each input location are computed. The sum of ECAL and HCAL over a given area is known as the transverse energy (ET). for(i = 1 to n) do for(j = 1 to m) do channel_et[i][j] = ECAL_filtered[i][j] +
View Full Document
Unlocking...