Another problem with lock free algorithms: sometimes locks are cheap. This is certainly the case on NVIDIA's CUDA platform, where atomic operations incur a 50%

Another problem with lock free algorithms: sometimes locks are cheap. This is certainly the case on NVIDIA's CUDA platform, where atomic operations incur a 50% penalty - which can sound like a lot but considering that the lock region can be 20% of runtime. In practice this means a lock free runs in 100 seconds while one with primitive locks runs in 110 seconds (and was easy to write debug and maintain).