Workload distribution for ray tracing in multi-core systems

Abstract

One of the features that made interactive ray tracing possible over the last few years was the careful exploitation of the computational power and parallelism available on modern multicore processors. Multithreaded interactive ray tracing engines have to share the workload (rays to be processed) among rendering threads. This may be achieved by storing tasks on a shared FIFO-queue, accessed by all threads. Accessing this shared data structure requires a data access control mechanism, which ensures that the data structure is not corrupted. This access mechanism must incur minimal overheads such that performance is not penalized. This paper proposes a lock-free data access control mechanism to such queue, which avoids all locks by carefully reordering instructions. This technique is compared with a classical lock-based approach and with a conservative local technique, where each thread maintains its local queue of tasks and shares nothing with other threads. Although the local approach outperforms the other two due to very good load balancing conditions, we demonstrate that the lock-free approach outperforms the lock-based one for large processor counts. Efficient and reliable sharing of data structures within a shared memory system is becoming a very relevant problem with the advent of many core processors. Lock free approaches are a promising manner of achieving such goal.