Re: When Parallel Doesn't Happen
- From: Kellyn Pedersen <kjped1313@xxxxxxxxx>
- To: Daniel Fink <daniel.fink@xxxxxxxxxxxxxx>
- Date: Thu, 1 Oct 2009 12:00:33 -0700 (PDT)
Hey Dan,
Now I could see raising parallel_max_servers if I was hitting the highwater or
lowering it if someone hadn't calculated the amount vs. number of CPU's, etc.
What I'm seeing is I have a setting of 168 for parallel_max_servers and yet a
total of 48 parallel servers are in use- when you look at the queries
involved, it should have allocated 96, (and at no time was it at the ceiling!)
parallel_automatic_tuning=false on these systems and the amount of parallel
servers allocated seem to vary depending on the load on the database. I am
working on a resource plan to put the higher priority jobs into a resource
group that would get "first dibs" on the available slaves, but I didn't
understand how dependent, (although it does make perfect sense to me in my
head...) parallel allocation is on resources. Not just CPU, but I/O,
especially when sequential reads are involved on indexes and although it shows
sequential reads for these- partitions.
I'm considering taking many of these from ASSM to manage the freelists
manually, giving the heavier hit partitioned tables a little more allocated
from the start...
Thoughts?
Kellyn
--- On Thu, 10/1/09, Daniel Fink <daniel.fink@xxxxxxxxxxxxxx> wrote:
From: Daniel Fink <daniel.fink@xxxxxxxxxxxxxx>
Subject: Re: When Parallel Doesn't Happen
To: kjped1313@xxxxxxxxx
Cc: oracle-l@xxxxxxxxxxxxx
Date: Thursday, October 1, 2009, 12:47 PM
Kellyn,
I've been working a similar problem. One 'working hypothesis' is that when a
session remains connected with an open cursor, the parallel server associated
with that session is not released. In this case, some of these sessions have
been idle for hours. Unfortunately, since we first came up with this hypothesis
and a workaround (temporarily raise parallel_max_servers for a stats gathering
job) was implemented, the problem has not come back. One friend, who is very
knowledgeable about PQ, has seen the same behaviour (he's Scottish...so the
spelling is correct for what he sees) coming from Toad.
PQ sounds like a great idea, but the actual user implementation and ability to
monitor what is actually happening is woefully inadequate. I do enjoy the
responses when I turn off parallelism and the full table scan with 1 process
runs significantly faster than the previous version with 8 slaves. One issue
that I often see is that a small degree of parallelism is used in dev/test with
only that process running. And it works just fine. But you take it to
production with 40 processes all runing a degree of 8 and things slow to a
crawl.
Regards,
Daniel Fink
--
Daniel Fink
OptimalDBA http://www.optimaldba.com
Oracle Blog http://optimaldba.blogspot.com
Lost Data? http://www.ora600.be/
Kellyn Pedersen wrote:
I have just started for a company that applies to the philosophy of too much of
a good thing is really a good thing, so bear with me...
The code has parallel hints everywhere- degrees often set to 8 in multiple
hints in one DML or DDL set, (parallel_threads_per_cpu=2, so start doing the
math...) The parallel_server is set anywhere from 96 to 168 and someone had
the idea that as long as they set the threshold high enough, everything would
run.
I have never seen parallel used the way it is here and I've come across some
very interesting challenges. The amount of slaves being allocated to a process
are being downgraded as resources are becoming limited by the poor choices made
in some of these simultaneous requests.
queries parallelized 10443
DDL statements parallelized 106
DFO trees parallelized 10549
Parallel operations not downgraded 10457
Parallel operations downgraded to serial 19
Parallel operations downgraded 75 to 99 pct 0
Parallel operations downgraded 50 to 75 pct 0
Parallel operations downgraded 25 to 50 pct 92
Parallel operations downgraded 1 to 25 pct 0
PX local messages sent 205011362
PX local messages recv'd 230002397
PX remote messages sent 0
PX remote messages recv'd 0
Other related posts:
