Re: memory mapped store entries

I was on the phone arguing with a `customer' when I typed this. I'd
better clarify.

On Thu, Aug 27, 1998 at 10:27:40AM +1200, Chris Wedgwood wrote:
> Since dirty pages probably won't be flushed by the kernel until
> they are paged out of explicitly msync'd - it should be more often,
> much more often.
>
> We don't have to sync the whole file either, we can get away with
> sync parts of it, and roll a window over the swap so that we sync
> it bit by bit.

Dirty pages won't normally be sync'd to disk by the kernel until they
are paged out - and when this is happening, your squids performance
will suck badly anyhow.

This means, the data has to be explicitly msync'd fairly often - but
we don't need to sync the entire swap at any one time, we can do it
bit by bit.

It also means, we can't always assume the swap is completely reliable
and consistent with the store, it might be out of date, in which
case, we have to rebuild part or all of it.

-cw
Received on Tue Jul 29 2003 - 13:15:53 MDT