Maurice Herlihy and Nir Shavit
understand the fear that keeps people in the high-tech industry awake
at night -- the fear that computers will become like washing machines.
"When you buy a washing machine you keep it until it breaks," Herlihy
says. "You don't trade it in every two years for a better model because
next year's washing machine isn't going to be significantly better than
this year's model."
So, are these two Sun researchers -- part-time members of the Scalable Synchronization Research Group in Sun Labs -- saying that Moore's Law has run its course? Are they saying that computers will no longer become twice as fast every year or every year and a half?
It is not they who say it, it has been widely recognized that we are up a against Moore's wall. Moore's law was never really about speed, it was about transistor density. We have reached the limit of what is easily obtainable (i.e. in Moore's 18 month period) in clock speed, and now the concentration is on adding cores. The problem is that adding cores does not increase your performance for scalar code, and for code that requires cross-cpu communications or locking, it can slow the performance down. That is what this paper is all about, reducing the locking and cross-cpu communications so that the performance will scale with the number of cores.
Posted by Brian Utterback on March 16, 2009 at 09:29 AM PDT #