64-bit size pgbench

Attached is a patch that fixes a long standing bug in pgbench: it won't
handle scale factors above ~4000 (around 60GB) because it uses 32-bit
integers for its computations related to the number of accounts, and it
just crashes badly when you exceed that. This month I've run into two
systems where that was barely enough to exceed physical RAM, so I'd
expect this to be a significant limiting factor during 9.0's lifetime.
A few people have complained about it already in 8.4.

The index size on the big accounts table has to increase for this to
work, it's a bigint now instead of an int. That's going to mean a drop
in results for some tests, just because less index will fit in RAM.
I'll quantify that better before submitting something final here. I
still have some other testing left to do as well: making sure I didn't
break the new \setshell feature (am suspicious of strtol()), confirming
the random numbers are still as random as they should be (there was a
little bug in the debugging code related to that, too).

Was looking for general feedback on whether the way I've converted this
to use 64 bit integers for the account numbers seems appropriate, and to
see if there's any objection to fixing this in general given the
potential downsides.

Here's the patch in action on previously unreachable sizes (this is a
system with 8GB of RAM, so I'm basically just testing seek speed here):

$ ./pgbench -j 4 -c 8 -T 30 -S pgbench
starting vacuum...end.
transaction type: SELECT only
scaling factor: 5000
query mode: simple
number of clients: 8
number of threads: 4
duration: 30 s
number of transactions actually processed: 2466
tps = 82.010509 (including connections establishing)
tps = 82.042946 (excluding connections establishing)

$ psql -x -c "select relname,reltuples from pg_class where
relname='pgbench_accounts'" -d pgbench
relname | pgbench_accounts
reltuples | 5e+08

$ psql -x -c "select pg_size_pretty(pg_table_size('pgbench_accounts'))"
-d pgbench
pg_size_pretty | 63 GB

$ psql -x -c "select aid from pgbench_accounts order by aid limit 1" -d
pgbench
aid | 1

$ psql -x -c "select aid from pgbench_accounts order by aid desc limit
1" -d pgbench
aid | 500000000

--
Greg Smith 2ndQuadrant Baltimore, MD
PostgreSQL Training, Services and Support
greg(at)2ndQuadrant(dot)com www.2ndQuadrant.com

Greg Smith <greg(at)2ndquadrant(dot)com> wrote:

> Attached is a patch that fixes a long standing bug in pgbench: it won't
> handle scale factors above ~4000 (around 60GB) because it uses 32-bit
> integers for its computations related to the number of accounts, and it
> just crashes badly when you exceed that. This month I've run into two
> systems where that was barely enough to exceed physical RAM, so I'd
> expect this to be a significant limiting factor during 9.0's lifetime.
> A few people have complained about it already in 8.4.

+1 for the fix.

Do we also need to adjust "tuples done" messages during dataload?
It would be too verbose for large scale factor. I think a message
every 1% is reasonable.

if (j % 10000 == 0)
fprintf(stderr, INT64_FORMAT " tuples done.\n", j);

Regards,
---
Takahiro Itagaki
NTT Open Source Software Center

Greg Smith <greg(at)2ndquadrant(dot)com> writes:
> Was looking for general feedback on whether the way I've converted this
> to use 64 bit integers for the account numbers seems appropriate, and to
> see if there's any objection to fixing this in general given the
> potential downsides.

In the past we've rejected proposed patches for pgbench on the grounds
that they would make results non-comparable to previous results. So the
key question here is how much this affects the speed. Please be sure to
test that on a 32-bit machine, not a 64-bit one.

> ! retval = (int64) strtol(res, &endptr, 19);

That bit is merely wishful thinking :-(

regards, tom lane

On Fri, Jan 29, 2010 at 11:09 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Greg Smith <greg(at)2ndquadrant(dot)com> writes:
>> Was looking for general feedback on whether the way I've converted this
>> to use 64 bit integers for the account numbers seems appropriate, and to
>> see if there's any objection to fixing this in general given the
>> potential downsides.
>
> In the past we've rejected proposed patches for pgbench on the grounds
> that they would make results non-comparable to previous results.

Perhaps we need an option indicating whether or not the use of bigint
columns is OK.

...Robert

Tom Lane wrote:
> In the past we've rejected proposed patches for pgbench on the grounds
> that they would make results non-comparable to previous results. So the
> key question here is how much this affects the speed. Please be sure to
> test that on a 32-bit machine, not a 64-bit one.
>

Sheesh, who has a 32-bit machine anymore? I'll see what older hardware
I can dig up. I've realized there are two separate issues to be
concerned about:

1) On small scale data sets, what's the impact of the main piece of data
being shuffled around in memory (the account number in the accounts
table) now being 64 bits? That part might be significantly worse on
32-bit hardware.

2) How does the expansion in size of the related primary key on that
data impact the breakpoint where the database doesn't fit in RAM anymore?

I did just updated my pgbench-tools package this month so that it
happily runs against either 8.3 or 8.4/9.0 and I've done two rounds of
extensive test runs lately, so plenty of data to compare against here.

>> ! retval = (int64) strtol(res, &endptr, 19);
>>
>
> That bit is merely wishful thinking :-(
>

I did specificially say I didn't trust that call one bit.

There is a middle ground position here, similar to what Robert
suggested, that I just add a "large mode" to the program for people who
need it without touching the current case. That might allow me to
sidestep some of these issues I may not have a good answer to with
getting the \setshell feature working right in 64 bits, could just make
that one specific to "regular mode".

In any case, I think this limitation in what pgbench can do has risen to
be a full-on bug at this point for the expected users of the next
version, and I'll sit on this until there's something better we can make
available.

--
Greg Smith 2ndQuadrant Baltimore, MD
PostgreSQL Training, Services and Support
greg(at)2ndQuadrant(dot)com www.2ndQuadrant.com