From: Kanoj S. <kan...@ya...> - 2002-04-05 21:14:06
|
. > Kanoj examined Dan's approach fairly aggressivly > especially with regard to TLB and cacheline True. > utilization. Dan was > able to convince him it was the right way over the > course of > dinner. Ummm, I seem to have given the wrong impression to Dan :-) I don't think it was quite clear to me exactly how/what Dan's patch improves over the current DISCONTIGMEM (possibly because I didn't understand the implementation), as well as how it will solve some of the problems associated with virtually contiguous but physically discontiguous chunks of mem_maps. I did mention that Dan should talk to Davis Mossberger to understand why this approach was tried and discarded on IA64 (other cpus might have different enough characterstics to make this approach performance safe). Some of the thought process that went into current DISCONTIGMEM is archived on oss.sgi.com/projects/numa (well, I hope it still is when oss comes back online!), but the set of restrictions that existed in the 2.3 timeframe might have changed, allowing improvements. I think DISCONTIGMEM should be looked as a feature that allows laying out of chunked mem_map arrays onto the chunks of physical memory they represent. There are possibly enough hooks to allow architectures to chunk up their available physical memory in any manner they wish (within reason). Agreed, this code might not be pretty, and could be improved. SGI folks, you should look closely at replacements of DISCONTIGMEM, since DISCONTIG is used by mips64/SN0 and ia64/SN1. Kanoj > __________________________________________________ Do You Yahoo!? Yahoo! Tax Center - online filing with TurboTax http://taxes.yahoo.com/ |