IDS Forum

Re: Question about sysindexes.clust

Posted By: Art Kagel
Date: Wednesday, 19 November 2008, at 10:55 a.m.

In Response To: Re: Question about sysindexes.clust (Eric Rowell)

Eric,

Thinking on this some more in my sleep, I realized that even though there
are some 69MM rows in the qty_det table I can't imagine why this query even
takes 4+ hours to complete at its best. Several items to consider going
forward:

- I remembered that there are 5 index accesses on the same index on the

qty_head and that all five are doing sub-key scans between the OR'd value

pairs of set_no and version. Breaking that OR into a UNION of two queries

will instead allow separate equality searches on those two keys along with

the two higher level keys which lead the index. The engine is doing this

internally to satisfy the IN clause on dept_code - hence the 5 independent

accesses to the index - but it cannot do so to satisfy the OR clause on two

columns. You'll have to resolve that yourself (you might even test

deconstructing this query into 10 UNIONED SELECT statements to resolve the

OR and the IN clauses into ten equality searches. Only testing will tell

which strategy is best. I always say there are at least three ways to

formulate ANY SELECT query. In your case there are at least eleven.

- No data is returned from qty_head so if that index also contained the

serial_key column access to this table would not require reading any data

pages and could be performed key-only rather than key-first. This simple

change would remove at least half and likely >90% of the IOs on this table.

Also, the added column may allow you to mark an otherwise duplicate index as

unique which generates a more efficient structure on disk. So, append

serial_key to this important join index.

- I see SERIAL FRAGMENT ALL for all three tables. Are the tables

fragmented? If not, perhaps they should be, certainly the two larger

tables. If they are, perhaps you should revisit the fragmentation scheme to

permit fragment elimination. Either way this query may benefit, as the

Clown suggested, from running under positive PDQPRIORITY to at least enable

fragment elimination and perhaps parallel searching of multiple fragments.

- What are the underlying disk structures that these IOs are so slow? In

4.5 hours you could read the entire qty_det table at a rate of only 99 IOs

per second! Any decent disk worth its salt should be able to sustain close

to 10 times that rate and that's not even considering the multiplier of any

multi-spindle RAID implementation under the hood. Something is NOT RIGHT

here.

- What are the underlying disk structures (single spindle, RAID5,

RAID1, RAID10, ???)

- Is there perhaps a problem with the writeback cache on the disk

array or SAN? A disabled cache can slow down a fast array by an order of

magnitude.

- Are you using RAW or COOKED chunks?

- Is the engine using KAIO for the chunk IO?

- Platform and OS and IDS versions?

Art

On Wed, Nov 19, 2008 at 9:28 AM, Eric Rowell <erowell@gmail.com> wrote:

> Art,
>
> Thank you for your analysis we came up with the same thing but backed
> into this via watching the page reads on the table during the applications
> run. We happened to find this as a production issue when trying to
> normalize this table in development so that the large highly repeated
> varchar coud be moved to another table.
>
> What was odd to me was that the Buffer Turn Over, Buffer Wait, and Read
> Ahead Utilization all doesn't appear to change for the worse, just the
> runtime and page reads (note we don't report on total page reads until now
> since day to day useage is different).
>
> I was embarrased at first about not monitoring this value but the more
> people tell me that they don't watch this value the more I understand it is
> not do to my own lack of trying. It appears that most people don't look at
> the sysindexes.clust value since like any go DB/SA data is moved around and
> reloaded for other reasons as a course of normal business. We too haven't
> run into this because we normally migration to new hardware every 3 years
> and about every 1 1/2 year find a reason to reload this table. At the end
> of last year we should have looked at reloading some tables since the
> migration to new hardware wasn't required. This would be way we haven't
> seen this before.
>
> It's now time for us to come up with some matrix to monitor this over
> time to see when this value indicates the need to reload or cluster a table
> just incase for the future. But it looks like a table by table a different
> value wouldn't indicate this need.
>
> Thanks again for all that replied,
>
> Eric B. Rowell
>

> <Previous entries SNIPPED>
>

--
Art S. Kagel
Oninit (www.oninit.com)
IIUG Board of Directors (art@iiug.org)

Disclaimer: Please keep in mind that my own opinions are my own opinions and
do not reflect on my employer, Oninit, the IIUG, nor any other organization
with which I am associated either explicitly or implicitly. Neither do
those opinions reflect those of other individuals affiliated with any entity
with which I am affiliated nor those of the entities themselves.

Messages In This Thread

• Question about sysindexes.clust
  Eric Rowell -- Tuesday, 18 November 2008, at 12:17 p.m.
  • Re: Question about sysindexes.clust
    Obnoxio The Clown -- Tuesday, 18 November 2008, at 12:22 p.m.
    • Re: Question about sysindexes.clust
      Eric Rowell -- Tuesday, 18 November 2008, at 12:45 p.m.
      • Re: Question about sysindexes.clust
        Obnoxio The Clown -- Tuesday, 18 November 2008, at 12:48 p.m.
        • Re: Question about sysindexes.clust
          Eric Rowell -- Tuesday, 18 November 2008, at 12:57 p.m.
          • Re: Question about sysindexes.clust
            Obnoxio The Clown -- Tuesday, 18 November 2008, at 1:00 p.m.
            • Re: Question about sysindexes.clust
              Obnoxio The Clown -- Tuesday, 18 November 2008, at 1:03 p.m.
              • Re: Question about sysindexes.clust
                Eric Rowell -- Tuesday, 18 November 2008, at 2:36 p.m.
                • Re: Question about sysindexes.clust
                  Obnoxio The Clown -- Tuesday, 18 November 2008, at 3:01 p.m.
                  • Re: Question about sysindexes.clust
                    Obnoxio The Clown -- Tuesday, 18 November 2008, at 3:04 p.m.
                    • Re: Question about sysindexes.clust
                      Eric Rowell -- Tuesday, 18 November 2008, at 3:28 p.m.
                      • Re: Question about sysindexes.clust
                        Obnoxio The Clown -- Tuesday, 18 November 2008, at 3:32 p.m.
                        • Re: Question about sysindexes.clust
                          Eric Rowell -- Tuesday, 18 November 2008, at 3:41 p.m.
                      • Re: Question about sysindexes.clust
                        Obnoxio The Clown -- Tuesday, 18 November 2008, at 3:33 p.m.
                        • Re: Question about sysindexes.clust
                          Eric Rowell -- Tuesday, 18 November 2008, at 3:42 p.m.
                          • Re: Question about sysindexes.clust
                            Obnoxio The Clown -- Tuesday, 18 November 2008, at 3:45 p.m.
                            • Re: Question about sysindexes.clust
                              Eric Rowell -- Tuesday, 18 November 2008, at 3:48 p.m.
                              • Re: Question about sysindexes.clust
                                Obnoxio The Clown -- Tuesday, 18 November 2008, at 3:54 p.m.
                                • Re: Question about sysindexes.clust
                                  Eric Rowell -- Tuesday, 18 November 2008, at 5:02 p.m.
                  • Re: Question about sysindexes.clust
                    Eric Rowell -- Tuesday, 18 November 2008, at 3:24 p.m.
                  • Re: Question about sysindexes.clust
                    Richard Snoke -- Tuesday, 18 November 2008, at 3:37 p.m.
                    • Re: Question about sysindexes.clust
                      Eric Rowell -- Tuesday, 18 November 2008, at 3:46 p.m.
                  • Re: Question about sysindexes.clust
                    Art Kagel -- Tuesday, 18 November 2008, at 4:28 p.m.
                    • Re: Question about sysindexes.clust
                      Obnoxio The Clown -- Tuesday, 18 November 2008, at 4:48 p.m.
                      • Re: Question about sysindexes.clust
                        Obnoxio The Clown -- Tuesday, 18 November 2008, at 4:52 p.m.
                        • Re: Question about sysindexes.clust
                          Eric Rowell -- Tuesday, 18 November 2008, at 5:06 p.m.
                          • Re: Question about sysindexes.clust
                            Obnoxio The Clown -- Tuesday, 18 November 2008, at 5:08 p.m.
                            • Re: Question about sysindexes.clust
                              Eric Rowell -- Wednesday, 19 November 2008, at 9:05 a.m.
                    • Re: Question about sysindexes.clust
                      Eric Rowell -- Wednesday, 19 November 2008, at 9:28 a.m.
                      • Re: Question about sysindexes.clust
                        Art Kagel -- Wednesday, 19 November 2008, at 10:55 a.m.
            • Re: Question about sysindexes.clust
              Eric Rowell -- Tuesday, 18 November 2008, at 2:18 p.m.
              • Re: Question about sysindexes.clust
                Obnoxio The Clown -- Tuesday, 18 November 2008, at 2:24 p.m.

IDS Forum is maintained by Administrator with WebBBS 5.12.