Gregory Stark wrote:
> "Gregory Stark" <stark@enterprisedb.com> writes:
>
> > "Bruce Momjian" <bruce@momjian.us> writes:
> >
> >> shared_buffers again was 32MB so all the data was in memory.
> >
> > The case where all the data is in memory is simply not interesting. The cost
> > of TOAST is the random access seeks it causes. You seem to be intentionally
> > avoiding testing the precise thing we're interested in.
>
> Also, something's not right with these results. 100,000 tuples --even if all
> they contain is a toast pointer-- won't fit on a single page. And the toast
> tables should vary in size depending on how many toast chunks are created.
The test creates _one_ row of length 100,000 and then finds out how long
it takes to access it twenty times.
I don't see how having the data outside cache helps us. For a large row
with 2k chunks, I assume all the 2k chunks are going to be in the same
8k page. What I want to measure is the cost of accessing four 2k chunks
vs. one 8k chunk, and I think we can conclude that is 6% of the access
time.
-- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB
http://www.enterprisedb.com
+ If your life is a hard drive, Christ can be your backup. +