I can't say anything specifically about this case, but generally, sometimes the whole architecture has to be changed in order to support concurrency (especially in the "millions of mutexes" kinds of problems). And such alternative architectures are often inherently inefficient in serial execution.