[Mulgara-dev] Transaction commit performance

[Paul Gearon]

On Jun 14, 2008, at 12:08 AM, Life is hard, and then you die wrote:

> On Fri, Jun 13, 2008 at 11:41:00PM -0500, Paul Gearon wrote:
>> This isn't actually an OS-dependent thing. It uses the TLB in the CPU
>> to know which pages are dirty.
>
> Sorry to nit pick, but that isn't quite accurate: dirty'ness is
> tracked in the page-tables; but in the sense that the TLB is a cache
> for the page-tables it does contain the dirty flags for those pages
> whose table entries currently happen to be in the TLB. However, on a
> flush the page-tables are scanned to figure out which entries are
> dirty, so the TLB doesn't come into play (directly) here (though when
> the data on the dirty pages is actually read their page-table entries
> will presuably get loaded into the TLB).

OK you're right, but I was glossing over this to get to my point. To  
interact with the TLB hardware there are established patterns that you  
have to screw up to get wrong. It's *possible* to screw it up, but I  
don't believe that even MS could have done that.

Paul