[Boost-users] boost::thread_specific_ptr performance?

Hi! What performance characteristics can be expected from a thread_specific_ptr on various platforms, as compared to other synchronizing primitives? Does accessing it cost the equivalent of locking a mutex, or is it generally faster than that? Specifically, would the cost of accessing it be noticeable if one were to use it (naively) to manage detailed per-thread logging, generating many thousands of access calls per second or more. Thanks in advance! Best regards, Josef Grahn