Mike D. Day wrote: > On 02/08/06 10:21 +0100, Keir Fraser wrote: > >> I wonder how it will interact with our late-pin/early-unpin model >> where we can directly write to pagetables before they are first used >> and also while they are being finally destroyed. It may be that if we >> use explicit batching and turn off that pinning logic we will not >> affect performance much, but we'll need to do some performance analysis. > > Where are we on Andrew's Patch now? I think Andrew's Patch is a good > solution for the time being until we can get batching, which may be > too disruptive for 3.03. > Mike I would agree :) Honestly, there's does not appear to be any evidence that batching of PTE updates, on typical applications, appear to have any performance advantage, but the current batching implementation does hurt scalability, and what really worries me, keeping it means keeping around a lot of code in xen mm that has to catch coherency issues (all the cleanup_writable_pagetable() calls and similar checks). I rather have writable PTs removed completely now (and assume they don't come back), then add explicit batching/multi-calls to specific parts of the guests' kernels (which should not require extra code in Xen mm like the coherency checks for writable PTs) -and only if they show to benefit from it. -Andrew
