Adrian Chadd <adr...@creative.net.au> writes: > On Wed, Oct 28, 2009, Alan Tyree wrote: > >> That's interesting, Daniel. So what are my tradeoffs. Run the normal kernel: >> faster but only 2.5gb; Run the bigmem kernel and suffer the performance but >> have more memory. > > Hm, does Linux actually -copy- data around when doing PAE?
Sometimes, the same way as the FreeBSD stuff you mention, using "bounce buffers" to help incapable hardware out. The big performance hit is that 32-bit needs to keep page tables in low memory, which is a scarce resource, and that it needs to use the kmap / kunmap infrastructure to provide access to pages in high memory, outside the easily accessed address space. That later is where most of the cost comes from. > The whole point behind PAE is to give you a larger amount of address space > usable by concurrent processes. So each process will have a differently > mapped set of "PAE" RAM pages as needed, along with the shared kernel stuff. Yeah. The real cost is the HIGHMEM stuff, not PAE directly, although that does increase pressure in low memory because the page tables are a bit bigger. Daniel -- ✣ Daniel Pittman ✉ dan...@rimspace.net ☎ +61 401 155 707 ♽ made with 100 percent post-consumer electrons Looking for work? Love Perl? In Melbourne, Australia? We are hiring. -- SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/ Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html