> On Jan 19, 2005, at 1:06 PM, Joakim Tjernlund wrote: > > > 8MB pages for user space? Isn't that a bit big? > > Just for kernel space. I think eliminating lots of TLB updates for the > kernel will be one of the easiest ways to get a boost of performance > in this area. This way, system calls or interrupts won't pollute the > TLB > for the application, which can then just continue to run without having > to reload the TLB after such events.
I see, similar to the pinned TLBs in 8xx but without the "pin" This would include vmalloc space and ioremapped space as well? > > Part of my graduate work years ago was MMU overhead with various > page sizes and replacement algorithms. Increase page sizes from > something like 1K to 8M made a difference you could measure, but > the difference between 1K to 32K made little difference. Of course, > you can always write some application that just trashes the TLB and > proves any improvement incorrect, but that would also trash caches and > other system resources. I'll always contend that if you can measure > the difference between a 4K and 16K page, something else is grossly > wrong with the system in general and you should be looking for > performance problems elsewhere. hmm, did you measure this on a 8xx with that have few TLBs? I havn't done any measurements, it just seemed like a good idea for systems with few TLBs and lots of RAM. Jocke