Thanks for the reference. Interesting. Looks like L1 cache is always private to the core. L2 and above may be shared among all cores or not. The cache logic can get quite complex. Looks like there is as many possibilities for performance optimization with cache as with multiple cores. But probably inaccessible from C and is so different between models that it is probably unusable except in the most intensive applications. And they are looking at doing memoization in cache.
On Tue, Feb 16, 2010 at 10:45 AM, David Mitchell <[email protected]>wrote: > On 2/16/2010 12:12, Don Guinn wrote: > > On Mon, Feb 15, 2010 at 7:42 PM, bill lam<[email protected]> wrote: > > > >> . . . Given the large memory cache available in modern > >> computer, it is better to just keep memory reference locality. The > >> speed of transfer of data using socket would be slower than that > >> within cache by at least 2 order of magnitude, just wild guess. . . . > >> > >> This touches on a question I have been wondering about. Is the cache > shared > > among multiple cores or does each one have its own cache? I would be > > surprised if it is shared as that could result in a big performance hit > if > > there are lots of cores. If the cache is private to each core then > somehow > > the data has to be moved from one cache to another. Main memory? > > ---------------------------------------------------------------------- > > For information about J forums see http://www.jsoftware.com/forums.htm > > > > http://en.wikipedia.org/wiki/L2_cache > > -- > David Mitchell > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm > ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
