[Prev] Thread [Next]  |  [Prev] Date [Next]

[coyotos-dev] Change to PATTs Jonathan S. Shapiro Thu Aug 31 03:58:39 2006

I'm not going to get the specification revised tonight, but we are going
to make a change to the PATT structure. The 'r' field will become
per-slot, and in consequence the PATT structure is now much closer to
the GPT design.

The reason for this change is that the "one r per patt" rule results in
a situation where there is no tree construction algorithm that does a
good job in all cases. Depending on the order in which tree inserts are
done, you may have to rebalance the tree, and this is very much beyond
anything we want to try in the kernel.

Making the 'r' value be per-slot adds the cost of a per-slot shift, but
this is not a significant factor in overall performance. The main factor
is that PATTs incur an extra cache miss in comparison to nodes.
Actually, the number of extra misses is very sensitive to the processor
family, and we need to look at this pretty carefully.

All in all, however, I think this is a negligable effect, and I continue
to think that PATTs are a better structure than nodes for our purposes;
even at the cost of the extra cache miss.


coyotos-dev mailing list