Re: Memory Management - BSD vs Linux

Cort (cort@cs.nmt.edu)
Tue, 12 Aug 1997 20:45:35 -0600 (MDT)


Why would the grouping of your accesses matter? The ppc cache line size
is 32b and the pte's (linux managed and hw managed) are 64b. Either way
you're going to be reloading a lot of cache.

It just so happens when you're accessing the page table tree other parts
of linux have likely touched it recently where with the ppc htab nothing
but the low-level page fault handlers mess with it.

I'm wondering how the hardware deals with a load of the tlb from the hash
table which is done automatically by the hw on the 604. Does that get
cached? I'll have to check the databook but that would effect things
quite a bit since the on-chip tlb is quite small and the hash table is
quite large (lots of accesses and thus cache loads).

}We were just talking about this and my guess was just the reverse.
}The hash tables are designed to scatter so that series of
}references to adjacent virtual page numbers should distribute over the
}table. The multi-level scheme localizes ptes. So, first guess is that
}bypassing hash table should make the dcache work better.
}Why is that wrong?