Speaker
Description
We present a study on latency and resource requirements for deep learning algorithms to run on a typical High Level Trigger computing farm at a high-pT LHC experiment at CERN. As a benchmark, we consider convolutional and graph autoencoders, developed to perform real-time anomaly detection on all the events entering the High Level Trigger (HLT) stage. The benchmark dataset consists of synthetic multijet events, simulated at a center-of-mass energy 13 TeV. Having in mind a next-generation heterogeneous computing farm powered with GPUs, we consider both optimized CPU and GPU inference, using hardware-specific optimization tools to meet the constraints of real-time processing at the LHC: ONNX runtime for CPU and NVIDIA TensorRT for GPUs. We observe O(msec) latency with different event batch sizes for both CPU- and GPU-based model inference with maximal gain seen at batch size of 1 (corresponding to the typical use case of event-parallelized HLT farms). We show that these optimized workflows offer significant savings with respect to native solutions (Tensorflow 2 and Keras) both in terms of time and computing resources.
