Re: [PERFORM] pgbench to the MAXINT

Attached is an updated 64-bit pgbench patch that works as expected for
all of the most common pgbench operations, including support for scales
above the previous boundary of just over 21,000. Here's the patched
version running against a 303GB database with a previously unavailable
scale factor:

$ pgbench -T 300 -j 2 -c 4 pgbench
starting vacuum...end.
transaction type: TPC-B (sort of)
scaling factor: 25000
query mode: simple
number of clients: 4
number of threads: 2
duration: 300 s
number of transactions actually processed: 21681
tps = 72.249999 (including connections establishing)
tps = 72.250610 (excluding connections establishing)

And some basic Q/A that the values it touched were in the right range:

$ psql -d pgbench -c "select min(aid),max(aid) from pgbench_accounts";

min | max
-----+------------
1 | 2500000000

$ psql -d pgbench -c "select min(aid),max(aid),count(*) from
pgbench_accounts where abalance!=0" &

min | max | count
-------+------------+-------
51091 | 2499989587 | 21678

(This system was doing 300MB/s on reads while executing that count, and
it still took 19 minutes)

The clever way Euler updated the patch, you don't pay for the larger
on-disk data (bigint columns) unless you use a range that requires it,
which greatly reduces the number of ways the test results can suffer
from this change. I felt the way that was coded was a bit more
complicated than it needed to be though, as it made where that switch
happened at get computed at runtime based on the true size of the
integers. I took that complexity out and just put a hard line in there
instead: if scale>=20000, you get bigints. That's not very different
from the real limit, and it made documenting when the switch happens
easy to write and to remember.

The main performance concern with this change was whether using int64
more internally for computations would slow things down on a 32-bit
system. I thought I'd test that on my few years old laptop. It turns
out that even though I've been running an i386 Linux on here, it's
actually a 64-bit CPU. (I think that it has a 32-bit install may be an
artifact of Adobe Flash install issues, sadly) So this may not be as
good of a test case as I'd hoped. Regardless, running a test aimed to
stress simple SELECTs, the thing I'd expect to suffer most from
additional CPU overhead, didn't show any difference in performance:

$ createdb pgbench
$ pgbench -i -s 10 pgbench
$ psql -c "show shared_buffers"
shared_buffers
----------------
256MB
(1 row)
$ pgbench -S -j 2 -c 4 -T 60 pgbench

i386 x86_64
6932 6924
6923 6926
6923 6922
6688 6772
6914 6791
6902 6916
6917 6909
6943 6837
6689 6744

6688 6744 min
6943 6926 max
6870 6860 average

Given the noise level of pgbench tests, I'm happy saying that is the
same speed. I suspect the real overhead in pgbench's processing relates
to how it is constantly parsing text to turn them into statements, and
that how big the integers it uses are is barley detectable over that.

So...where does that leave this patch? I feel that pgbench will become
less relevant very quickly in 9.1 unless something like this is
committed. And there don't seem to be significant downsides to this in
terms of performance. There are however a few rough points left in here
that might raise concern:

1) A look into the expected range of the rand() function suggests the
glibc implementation normally proves 30 bits of resolution, so about 1
billion numbers. You'll have >1B rows in a pgbench database once the
scale goes over 10,000. So without a major overhaul of how random
number generation is treated here, people can expect the distribution of
rows touched by a test run to get less even once the database scale gets
very large. I added another warning paragraph to the end of the docs in
this update to mention this. Long-term, I suspect we may need to adopt
a superior 64-bit RNG approach, something like a Mersenne Twister
perhaps. That's a bit more than can be chewed on during 9.1 development
though.

2) I'd rate odds are good there's one or more corner-case bugs in
\setrandom or \setshell I haven't found yet, just from the way that code
was converted. Those have some changes I haven't specifically tested
exhaustively yet. I don't see any issues when running the most common
two pgbench tests, but that's doesn't mean every part of that 32 -> 64
bit conversion was done correctly.

Given how I use pgbench, for data generation and rough load testing, I'd
say neither of those concerns outweights the need to expand the size
range of this program. I would be happy to see this go in, followed by
some alpha and beta testing aimed to see if any of the rough spots I'm
concerned about actually appear. Unfortunately I can't fit all of those
tests in right now, as throwing around one of these 300GB data sets is
painful--when you're only getting 72 TPS, looking for large scale
patterns in the transactions takes a long time to do. For example, if I
really wanted a good read on how bad the data distribution skew due to
small random range is, I'd need to let some things run for a week just
for a first pass.

I'd like to see this go in, but the problems I've spotted are such that
I would completely understand this being considered not ready by
others. Just having this patch available here is a very useful step
forward in my mind, because now people can always just grab it and do a
custom build if they run into a larger system.

Wavering between Returned with Feedback and Ready for Committer here.
Thoughts?

--
Greg Smith 2ndQuadrant US greg(at)2ndQuadrant(dot)com Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books