Description

The entire CMS event reconstruction is written in C++, configured in python enabling the use of dynamic data structures with high granularity. This feature is very useful because typical High-Energy Physics algorithms are fed with variable size inputs and produce outputs, whose size is many times impossible to estimate or generalize. The CMS software framework is going through a transition aimed at transforming it into a heterogeneous software framework, in order to profit of GPUs higher throughput and energy efficiency. CUDA language is used to program NVIDIA GPUs and it is not dynamic-data-structures-friendly, since it was designed for working with data structures of predictable size. Furthermore, allocating memory on a GPU has a fixed cost which is orders of magnitude higher than doing that on the host.

Task ideas

• Creation of a testbed for the creation and validation of STL-like data structures
• Design and implementation of arena allocators to allocate/resize/access a CUDA vector-of-vector
• Implementation of GPU-friendly tree-like data structure
• Smart allocators to preallocate memory and assign it to requesting functions, keeping track of the memory usage and reassigning it on demand.

Expected results

Creation of a library containing the developed data structures

Requirements

C++, CUDA

Mentors

• Felice Pantaleo
• Vincenzo Innocente
• Marco Rovere
• Andrea Bocci

Corresponding Project

• CMS

Participating Organizations

• CERN

---