Performance Metrics of a GPU Based Track Fitting Code for the INO Prototype Stack

The use of Graphical Processing Units (GPU) for achieving data level parallelism (DLP) in high energy physics computations is gaining attention. India-based neutrino observatory (INO) is a proposed experiment to study neutrinos. There is a wide scope for use of GPUs in the data acquisition system and data analysis for the Iron Calorimeter (ICAL) detector in INO. In this context, we explore the possibility of using GPUs for track fitting of the INO prototype stack data. We analyze the performance metrics using a linear fitting kernel with a dummy dataset. We also present the impact of memory bandwidth on the overall performance of the code. This will also be a guideline for other activities of INO like simulation, where application of a massively parallel architecture is envisaged.