Gregory Stark wrote:
> "Bruce Momjian" <bruce@momjian.us> writes:
>
> > Gregory Stark wrote:
> >> "Bruce Momjian" <bruce@momjian.us> writes:
> >>
> >> > I tested TOAST using a method similar to the above method against CVS
> >> > HEAD, with default shared_buffers = 32MB and no assert()s. I created
> >> > backends with power-of-2 seetings for TOAST_TUPLES_PER_PAGE (4(default),
> >> > 8, 16, 32, 64) which gives TOAST/non-TOAST breakpoints of 2k(default),
> >> > 1k, 512, 256, and 128, roughly.
> >> >
> >> > The results are here:
> >> >
> >> > http://momjian.us/expire/TOAST/
> >> >
> >> > Strangely, 128 bytes seems to be the break-even point for TOAST and
> >> > non-TOAST, even for sequential scans of the entire heap touching all
> >> > long row values. I am somewhat confused why TOAST has faster access
> >> > than inline heap data.
>
> Is your database initialized with C locale? If so then length(text) is
> optimized to not have to detoast:
>
> if (pg_database_encoding_max_length() == 1)
> PG_RETURN_INT32(toast_raw_datum_size(str) - VARHDRSZ);
Wow, we optimized length(). OK, will run the test with
substring(t,1,1).
> Also, I think you have to run this for small datasets like you have well as
> large data sets where the random access seek time of TOAST will really hurt.
Well, if everything doesn't fit in the cache, then the smaller heap will
be a bigger win for non-TOAST access, so some of that overhead balances
out. Let me get in-cache numbers and then I can look at larger data sets.
-- 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. +