Re: Cost limited statements RFC

On 6/6/13 4:02 PM, Robert Haas wrote:
> If we can see our way clear to ripping out the autovacuum costing
> stuff and replacing them with a read rate limit and a dirty rate
> limit, I'd be in favor of that.  The current system limits the linear
> combination of those with user-specified coefficients, which is more
> powerful but less intuitive.

There is also an implied memory bandwidth limit via the costing for a
hit, which was the constraint keeping me from just going this way last
time this came up.  It essentially limits vacuum to 78MB/s of scanning
memory even when there's no disk I/O involved.  I wasn't sure if that
was still important, you can also control it now via these coefficients,
and most of the useful disk rate refactorings simplify a lot if that's
gone.  The rest of this message is some evidence that's not worth
keeping though, which leads into a much cleaner plan than I tried to
pitch before.

I can now tell you that a busy server with decent memory can easily chug
through 7.8GB/s of activity against shared_buffers, making the existing
78MB/s limit is a pretty tight one.  7.8GB/s of memory access is 1M
buffers/second as measured by pg_stat_database.blks_read.  I've attached
a sample showing the highest rate I've seen as evidence of how fast
servers can really go now, from a mainstream 24 Intel cores in 2 sockets
system.  Nice hardware, but by no means exotic stuff.  And I can hit
500M buffers/s = 4GB/s of memory even with my laptop.

I have also subjected some busy sites to a field test here since the
original discussion, to try and nail down if this is really necessary.
So far I haven't gotten any objections, and I've seen one serious
improvement, after setting vacuum_cost_page_hit to 0.  The much improved
server is the one I'm showing here.  When a page hit doesn't cost
anything, the new limiter on how fast vacuum can churn through a well
cached relation usually becomes the CPU speed of a single core.
Nowadays, you can peg any single core like that and still not disrupt
the whole server.

If the page hit limit goes away, the user with a single core server who
used to having autovacuum only pillage shared_buffers at 78MB/s might
complain that if it became unbounded.  I'm not scared of that impacting
any sort of mainstream hardware from the last few years though.  I think
you'd have to be targeting PostgreSQL on embedded or weak mobile chips
to even notice the vacuum page hit rate here in 2013.  And even if your
database is all in shared_buffers so it's possible to chug through it
non-stop, you're way more likely to suffer from an excess dirty page
write rate than this.

Buying that it's OK to scrap the hit limit leads toward a simple to code
implementation of read/write rate limits implemented like this:

-vacuum_cost_page_* are removed as external GUCs.  Maybe the internal
accounting for them stays the same for now, just to keep the number of
changes happening at once easier.

-vacuum_cost_delay becomes an internal parameter fixed at 20ms.  That's
worked out OK in the field, there's not a lot of value to a higher
setting, and lower settings are impractical due to the effective 10ms
lower limit on sleeping some systems have.

-vacuum_cost_limit goes away as an external GUC, and instead the actual
cost limit becomes an internal value computed from the other parameters.
  At the default values the value that pops out will still be close to
200.  Not messing with that will keep all of the autovacuum worker cost
splitting logic functional.

-New vacuum_read_limit and vacuum_write_limit are added as a kB value
for the per second maximum rate.  -1 means unlimited.  The pair replaces
changing the cost delay as the parameters that turns cost limiting on or
off.

That's 5 GUCs with complicated setting logic removed, replaced by 2
simple knobs, plus some churn in the autovacuum_* versions.  Backwards
compatibility for tuned systems will be shot.  My position is that
anyone smart enough to have navigated the existing mess of these
settings and done something useful with them will happily take having
their custom tuning go away, if it's in return for the simplification.
At this point I feel exactly the same way I did about the parameters
removed by the BGW auto-tuning stuff that went away in 8.3, with zero
missing the old knobs that I heard.  Another year of experiments and
feedback has convinced me nobody is setting this usefully in the field
who wouldn't prefer the new interface.

--
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.com