Re: When Parallel Doesn't Happen

From: Daniel Fink <daniel.fink_at_optimaldba.com>
Date: Thu, 01 Oct 2009 12:47:16 -0600
Message-ID: <4AC4F934.6060808_at_optimaldba.com>



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
>
>
>
-- http://www.freelists.org/webpage/oracle-l
Received on Thu Oct 01 2009 - 13:47:16 CDT

Original text of this message