Re: WIP patch: reducing overhead for repeat de-TOASTing

Tom Lane wrote:

> OK, I've reproduced the test case locally. I believe that when you
> say "worse", you mean "worse than 8.3", right? And you did tell me
> offlist that you were testing with --enable-cassert. CVS HEAD has
> very substantially greater cassert overhead because of the
> randomize_memory addition --- oprofile output for this test looks like
>
> samples % image name symbol name
> 1239580 78.7721 postgres randomize_mem
> 143544 9.1218 libc-2.7.so memcpy
> 48039 3.0528 libc-2.7.so memset
> 13838 0.8794 postgres LWLockAcquire
> 12176 0.7738 postgres index_getnext
> 11697 0.7433 postgres LWLockRelease
> 10406 0.6613 postgres hash_search_with_hash_value
> 4739 0.3012 postgres toast_fetch_datum
> 4099 0.2605 postgres _bt_checkkeys
> 3905 0.2482 postgres AllocSetAlloc
> 3751 0.2384 postgres PinBuffer
> 3545 0.2253 postgres UnpinBuffer
>
> I'm inclined to think that we'd better turn that off by default,
> since it's not looking like it's catching anything new.

Yes, I suspect that's probably it. I applied the patch straight to CVS
tip as I wasn't aware of any changes that would affect the unpatched
result, but I was obviously wrong ;)

(cut)

> On the whole I'm still feeling pretty discouraged about this patch ...

At the very least we have some more information on how an eventual
solution should work, and a test case to help analyse the effectiveness
of any potential solution.

ATB,

Mark.

--
Mark Cave-Ayland
Sirius Corporation - The Open Source Experts
http://www.siriusit.co.uk
T: +44 870 608 0063