Re: BBU Cache vs. spindles - Mailing list pgsql-performance

Scott Marlowe wrote:
> On Wed, Oct 20, 2010 at 8:25 PM, Joshua D. Drake <jd@commandprompt.com> wrote:
> > On Wed, 2010-10-20 at 22:13 -0400, Bruce Momjian wrote:
> >> Ben Chobot wrote:
> >> > On Oct 7, 2010, at 4:38 PM, Steve Crawford wrote:
> >> >
> >> > > I'm weighing options for a new server. In addition to PostgreSQL, this machine will handle some modest Samba
andRsync load. 
> >> > >
> >> > > I will have enough RAM so the virtually all disk-read activity will be cached. The average PostgreSQL read
activitywill be modest - a mix of single-record and fairly large (reporting) result-sets. Writes will be modest as well
butwill come in brief (1-5 second) bursts of individual inserts. The rate of insert requests will hit 100-200/second
forthose brief bursts. 
> >> > >
> >> > > So...
> >> > >
> >> > > Am I likely to be better off putting $$$ toward battery-backup on the RAID or toward adding a second RAID-set
andsplitting off the WAL traffic? Or something else? 
> >> >
> >> > A BBU is, what, $100 or so? Adding one seems a no-brainer to me.
> >> > Dedicated WAL spindles are nice and all, but they're still spinning
> >> > media. Raid card cache is waaaay faster, and while it's best at bursty
> >> > writes, it sounds like bursty writes are precisely what you have.
> >>
> >> Totally agree!
> >
> > BBU first, more spindles second.
>
> Agreed.  note that while you can get incredible burst performance from
> a battery backed cache, due to both caching and writing out of order,
> once the throughput begins to saturate at the speed of the disk array,
> the bbu cache is now only re-ordering really, as it will eventually
> fill up faster than the disks can take the writes, and you'll settle
> in at some percentage of your max tps you get for a short benchmark
> run.  It's vitally important that once you put a BBU cache in place,
> you run a very long running transactional test (pgbench is a simple
> one to start with) that floods the io subsystem so you see what you're
> average throughput is with the WAL and data store getting flooded.  I
> know on my system pgbench runs of a few minutes can be 3 or 4 times
> faster than runs that last for the better part of an hour.

With a BBU you can turn off full_page_writes, which should decrease the
WAL traffic.

However, I don't see this mentioned in our documentation.  Should I add
it?

--
  Bruce Momjian  <bruce@momjian.us>        http://momjian.us
  EnterpriseDB                             http://enterprisedb.com

  + It's impossible for everything to be true. +