On Wed, Oct 01, 2003 at 03:37:48 -0000, sushil mayengbam wrote: > hi folks/ > i've some trouble in digesting the idea of Bhowmick's "slab > coloring" concept. > can anyone of u chew this things out taking some logical > example???. > i apologize if my question is too naive. > (and lastly, i must thank "Riek and mel Gorman " for their > valuable documents on MM.) I don't understand the hardware-related implications too well, but it's along the lines of: To make life easier, hardware caches are organized in cache-lines. Each cache-line always caches blocks (some eg 32 bytes wide) with same low address bits (above the cache-line width). So if the structure always started at, say 64-byte, boundary, it's first half (asume it's where the most used data are) would always use even cacheline. So the computer would soon ran out of even cachelines, while the odd ones would be unused. Thus the starting address for data in each page is shifted by some amount in each page. Thus the hot data will end up in different cache-lines for different instances of given objects. That leads to better cache utilization and thus speed up of operations like list walking. Note, that since last few bytes (3 pointers IIRC) of each page is used by slab cache itself, power-of-two sized structures, which most benefit from coloring, never fit the slab exactly, leaving good possibility for coloring. Then there is a lot of black magic involved in tweaking various constants for best performance on given architecture. ------------------------------------------------------------------------------- Jan 'Bulb' Hudec <bulb@ucw.cz> -- Kernelnewbies: Help each other learn about the Linux kernel. Archive: http://mail.nl.linux.org/kernelnewbies/ FAQ: http://kernelnewbies.org/faq/
