Research conducted by Goethe University Frankfurt.

Our research focused on event reconstruction in the CBM experiment, exploring parallel processing potential in other FIDIUM projects. The following conclusions have been made:

• Optimization. Programs and algorithms must be optimized for high computing performance.
• Proper hardware utilization. Proper hardware feature utilization can significantly speed up calculations, though not all tasks benefit equally.

For effective parallelization, the following requirements should be met:

• Suitable data structures. Suitable data structures should be used. Memory access should be sequential. For better performance, data structures should be reformatted beforehand.
• Cache utilization. Caches should be utilized to optimize memory and time usage. Sequential operations should be performed on a variable wherever possible to reduce the number of times the memory is accessed.
• Independent or synchronized variables in threads. Ensure variables of concurrently executing threads are independent or processes are clearly synchronized. GPU threads of the same core should execute the same instructions. Minimize or avoid branching and limit kernel size to the instruction cache size.

At this stage:

• The performance of representative reconstruction algorithms on CPU and GPU has been evaluated.
• Effects and applicability of data structures and parallelization models have been studied.
• Currently, the development of a generalized vector (Vc) and scalar version of the CBM L1 tracking algorithm [AP-III] is in progress.

Future work

Parallelization of CBM track reconstruction on GPU based on the previous generalized version, containerization, and ensuring efficient operation of the common CPU/GPU version of the track reconstructor on third-party computing clusters is planned for the future.