On Fri, Dec 22, 2023 at 8:14 AM Andrey M. Borodin <x4mmm@yandex-team.ru> wrote:
> Just a side node.
> It seems like commit log is kind of antipattern of data contention. Even when we build a super-optimized SLRU. Nearby
**bits**are written by different CPUs.
> I think that banks and locks are good thing. But also we could reorganize log so that
> status of transaction 0 is on a page 0 at bit offset 0
> status of transaction 1 is on a page 1 at bit offset 0
> status of transaction 2 is on a page 2 at bit offset 0
> status of transaction 3 is on a page 3 at bit offset 0
> status of transaction 4 is on a page 0 at bit offset 2
> status of transaction 5 is on a page 1 at bit offset 2
> status of transaction 6 is on a page 2 at bit offset 2
> status of transaction 7 is on a page 3 at bit offset 2
> etc...
This is an interesting idea. A variant would be to stripe across
cachelines within the same page rather than across pages. If we do
stripe across pages as proposed here, we'd probably want to rethink
the way the SLRU is extended -- page at a time wouldn't really make
sense, but preallocating an entire file might.
> And it would be even better if page for transaction statuses would be determined by backend id somehow. Or at least
cacheline. Can we allocate a range (sizeof(cacheline)) of xids\subxids\multixacts\whatever for each backend?
I don't understand how this could work. We need to be able to look up
transaction status by XID, not backend ID.
--
Robert Haas
EDB: http://www.enterprisedb.com