switch to thread-based parallelism

the previous process-based implementation can crash due to no pooling