Collaborative Research : Elements : Extending the physics reach of LHCb by developing and deploying algorithms for a fully GPU-based first trigger stage

粒子物理学标准模型（SM）的发展是一项重大的智力成就。在欧洲核子研究中心的大型强子对撞机（LHC）上发现的希格斯玻色子进一步证实了这一模型的有效性。然而，标准模型留下了许多悬而未决的问题，包括为什么物质在宇宙中占主导地位，以及暗物质的性质。大多数解释都需要新的现象，我们称之为超越标准模型物理学（BSM），而欧洲核子研究中心（CERN）的大型强子对撞机实验就是为了探索这种现象而设计的。大型强子对撞机是世界上首屈一指的高能物理粒子加速器，目前在瑞士日内瓦附近的欧洲核子研究中心实验室运行，该实验室是解决这些BSM问题的最重要设施之一。LHCb实验是LHC的四个大型实验之一，旨在详细研究含有b或c夸克的强子的衰变。目标是通过检查包含这些夸克的强子的性质，来确定标准模型之外的新物理的存在。新的物理，或者新的力，可以通过粒子来体现，但还没有被发现，它们的存在将改变含有b和c夸克的强子的衰变速率和CP违反不对称性，从而允许间接观察新现象-或者通过直接观察新的携带力的粒子。LHC实验收集的数据集是世界上最大的数据集之一。例如，两个pi都参与的LHCb实验的传感器阵列每年产生约100 TB/s，接近1 zb。即使在使用定制的读出电子设备大幅减少数据量之后，每年的数据量仍约为10艾字节。如此庞大的数据集不可能无限期地存储；因此，所有高能物理（HEP）实验都采用由数据摄取系统（在HEP中称为触发系统）实时执行的第二种数据缩减方案，以决定每个事件是要保留以供将来分析还是永久丢弃。该项目的主要目标是开发和部署软件，以最大限度地提高LHCb触发系统的性能-在gpu上运行其第一个处理阶段-从而实现LHCb的全部物理发现潜力。LHCb探测器正在为运行3进行升级（将于2022年开始记录数据），届时触发系统将需要每年处理25eb。目前，触发器每年处理的10艾字节中只有0.3个使用高级计算算法进行分析；在此阶段之前，使用fpga上执行的简单算法丢弃其余部分。为了在Run 3中显著扩展其物理覆盖范围，LHCb计划使用高级计算算法每年处理全部25艾字节。pi建议在gpu上运行整个第一个触发处理阶段，这具有零（可能为负）净成本，并为第二个处理阶段释放所有CPU资源。LHCb触发器大量使用机器学习（ML）算法，这些算法需要针对Run 3条件和gpu的使用进行重新优化。该提案的具体目标是：主要触发器选择算法的基于gpu的版本，它大量使用机器学习；基于gpu的热量聚类和电子识别算法，可能使用ML；以及在基于gpu的触发器框架内部署ML算法所需的基础设施。这些进步将使探索暗物质的许多潜在解释成为可能，例如，暗光子衰变，以及我们宇宙的物质/反物质不对称，这些数据将由于触发系统的限制而无法使用。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The development of the Standard Model (SM) of particle physics is a major intellectual achievement. The validity of this model was further confirmed by the discovery of the Higgs boson at the Large Hadron Collider (LHC) at CERN. However, the Standard Model leaves open many questions, including why matter dominates over anti-matter in the Universe and the properties of dark matter. Most explanations require new phenomena, which we call Beyond the Standard Model Physics (BSM), and which the LHCb experiment at CERN has been designed to explore. The LHC is the premier High Energy Physics particle accelerator in the world and is currently operating at the CERN laboratory near Geneva Switzerland, one of the foremost facilities for addressing these BSM questions. The LHCb experiment is one of four large experiments at the LHC and is designed to study in detail the decays of hadrons containing b or c quarks. The goal is to identify the existence of new physics beyond the Standard Model by examining the properties of hadrons containing these quarks. The new physics, or new forces, can be manifest by particles, as yet to be discovered, whose presence would modify decay rates and CP violating asymmetries of hadrons containing the b and c quarks, allowing new phenomena to be observed indirectly - or via direct observation of new force-carrying particles. The data sets collected by the LHC experiments are some of the largest in the world. For example, the sensor arrays of the LHCb experiment, in which both PIs participate, produce about 100 TB/s and close to a zettabyte per year. Even after drastic data-reduction performed by custom-built read-out electronics, the data volume is still about 10 exabytes per year. Such large data sets cannot be stored indefinitely; therefore, all high energy physics (HEP) experiments employ a second data-reduction scheme executed in real time by a data-ingestion system - referred to as a trigger system in HEP - to decide whether each event is to be persisted for future analysis or permanently discarded. The primary goal of this project is developing and deploying software that will maximize the performance of the LHCb trigger system - running its first processing stage on GPUs - so that the full physics discovery potential of LHCb is realized.The LHCb detector is being upgraded for Run 3 (which will start to record data in 2022), when the trigger system will need to process 25 exabytes per year. Currently, only 0.3 of the 10 exabytes per year processed by the trigger is analyzed using high-level computing algorithms; the rest is discarded prior to this stage using simple algorithms executed on FPGAs. To significantly extend its physics reach in Run 3, LHCb plans to process the entire 25 exabytes each year using high-level computing algorithms. The PIs propose running the entire first trigger-processing stage on GPUs, which has zero (likely negative) net cost, and frees up all of the CPU resources for the second processing stage. The LHCb trigger makes heavy use of machine learning (ML) algorithms, which will need to be reoptimized both for Run 3 conditions but also for usage on GPUs. The specific objectives of this proposal are developing: GPU-based versions of the primary trigger-selection algorithms, which make heavy usage of ML; GPU-based calorimeter-clustering and electron-identification algorithms, likely using ML; and the infrastructure required to deploy ML algorithms within the GPU-based trigger framework. These advances will make it possible to explore many potential explanations for dark matter, e.g., dark photon decays, and the matter/anti-matter asymmetry of our universe using data that would be otherwise inaccessible due to trigger-system limitations.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.