Snapshot Reuse

In Read Committed transactions we take snapshots much more frequently
than transactions begin and commit. It would be help scalability if we
didn't need to re-take a snapshot. That's only helpful if the chances of
seeing the snapshot is relatively high.

Now that we have virtual transactions we may more frequently find
ourselves taking identical snapshots.

If we had a counter that incremented each time the main snapshot altered
in a meaningful way we could set that atomically. We could then read
this when we take a snapshot to see if it matches our existing snapshot;
if so then drop the lock quickly and continue with what we already have.

I can see some downsides to this as well as potential benefits:

* we ping the counter across CPUs - yes, we will, but that's probably
better than pinging the whole procarray

* this relies upon the rate of change of snapshots - need to do the math
to see how often this might apply

Not sure yet myself, but it seems worth recording in case it spurs an
idea from someone else.

--
Simon Riggs
2ndQuadrant http://www.2ndQuadrant.com

I certainly would be interested to see if it improves performance.

---------------------------------------------------------------------------

Simon Riggs wrote:
> In Read Committed transactions we take snapshots much more frequently
> than transactions begin and commit. It would be help scalability if we
> didn't need to re-take a snapshot. That's only helpful if the chances of
> seeing the snapshot is relatively high.
>
> Now that we have virtual transactions we may more frequently find
> ourselves taking identical snapshots.
>
> If we had a counter that incremented each time the main snapshot altered
> in a meaningful way we could set that atomically. We could then read
> this when we take a snapshot to see if it matches our existing snapshot;
> if so then drop the lock quickly and continue with what we already have.
>
> I can see some downsides to this as well as potential benefits:
>
> * we ping the counter across CPUs - yes, we will, but that's probably
> better than pinging the whole procarray
>
> * this relies upon the rate of change of snapshots - need to do the math
> to see how often this might apply
>
> Not sure yet myself, but it seems worth recording in case it spurs an
> idea from someone else.
>
> --
> Simon Riggs
> 2ndQuadrant http://www.2ndQuadrant.com
>
>
> ---------------------------(end of broadcast)---------------------------
> TIP 2: Don't 'kill -9' the postmaster

--
Bruce Momjian <bruce(at)momjian(dot)us> http://momjian.us
EnterpriseDB http://postgres.enterprisedb.com

+ If your life is a hard drive, Christ can be your backup. +