You are viewing paulmck

(no subject)

inside
Something like this was tried in the 1990s, but the software plus caches out-performed the hardware "acceleration". It still takes time to propagate the information about who holds what lock across the system, and even though CPU core clock rates are limited at roughly 2003 levels, electrons just don't move fast enough.

The thing is that the key performance benefit of transactional lock elision is not a more-efficient implementations of locks, but instead omitting the locks altogether. Because the lock state need not be communicated (if things work out nicely), it is not necessary to wait for the electrons to do the communication. On the other hand, all the transactional-memory limitations apply, and even in the absence of these limitations, from what I can see, transactional lock elision results in some subtle changes in semantics. But more on that later...


Comment Form

No HTML allowed in subject

  
 
   
 

Notice! This user has turned on the option that logs your IP address when posting. 

(will be screened)