Just curious, how long does the relevant full table scan take on a single node,
serial, at low parallelism, and at high parallelism?
Possibly the full table scan (to dev/null) without any complications of the
rest of the query will avoid whatever complicated problem you are triggering
(which at the moment seems like an RBDMS deficiency, but I might not be seeing
something.)
Or, if the unadorned FTS also shows the problem intermittently the diagnostics
might be quite a bit simpler. (I am not proposing this as a long term solution.)
Since you have the testing harness in place it seems to me slapping this extra
step in place as a precursor to the actual query might show whether or not the
problem is an artifact of the query or a problem with the FTS apparatus itself.
Good luck LS!
mwf
From: oracle-l-bounce@xxxxxxxxxxxxx [mailto:oracle-l-bounce@xxxxxxxxxxxxx] On ;
Behalf Of Ls Cheng
Sent: Wednesday, June 22, 2022 1:40 AM
To: Tanel Poder
Cc: Oracle Mailinglist
Subject: Re: Parallel Query change from smart scan to cell single block access
Hi
There might be long transactions but since this system has Goldengate running
any transactions which last more than 2 hours are alarmed, if not exiting from
application are killed. But the query when the problem appears the PQ query
elapsed time increases to many hours to finish (as long as 11 hours in my
tests). In my tests I am executing 10 concurrent executions every 5 minutes and
during 12 hours tests I can reproduce the issue 3 or 4 times, the tests should
clean table blocks because every 5 minutes the query is executed and does Full
Table Scan.
Thanks
On Sun, Jun 19, 2022 at 7:47 AM Tanel Poder <tanel@xxxxxxxxxxxxxx> wrote:
That 60 second snapshot indicates that the full table scan / smart scan was not
making any progress during that time. No increases in "table scan blocks
gotten" nor "cell physical IO bytes eligible for predicate offload" or any
other "cell blocks%" metrics.
The "consistent gets", "consistent changes" and "transaction tables consistent
reads - undo records applied" have all increased by the same amount (68471)
without any table/index block access metrics. You might be stuck processing the
same block and looks like you might be hitting something like a delayed block
cleanout issue (no metric saying undo blocks applied to a datablock, just
"transaction tables consistent reads - undo records applied").
Smart scan can't process the block in the storage layer as it doesn't know
whether some previous changes to that block had been committed before or after
your smart scan started (snapshot SCN). So it sends the block back in
passthrough mode and Oracle now starts examining (and rolling back) a CR copy
of the undo segment header block, to get to the truth.
Thoughts:
1. Do you have any (unrelated) long running transactions active at the
same time? There's something called the "minimum active SCN" optimization where
Oracle keeps track of the lowest SCN of any still active (uncommitted)
transaction and can push this info to storage cells too. If a select query does
a consistent read on a block, requesting a snapshot SCN higher than the known
"minimum active SCN" in the database, then we can be sure that the still locked
rows (by a transaction with a SCN lower than the known minact SCN) are real and
no need to go examining & rolling back undo segment header blocks to a past
point in time.
2. You could also try to clean table blocks after loads or large ETL/batch
modifications by committing more often (so that the fast cleanout can clean the
blocks still in buffer cache on commit or just run a buffered (not direct path
or smart) full table scan on the table after data load).
3. Of course, if it's some bug, all bets are off, might be something more
complicated :-)
This is somewhat related to the topic (long running transactions causing some
optimizations not be usable):
* EXADATA and SuperCluster : Check if long running transactions are
preventing min active scn from progressing, resulting in Storage Indexes not
being used (Doc ID 2081483.1)
--
Tanel Poder
https://learn.tanelpoder.com
On Sat, Jun 18, 2022 at 8:10 PM Ls Cheng <exriscer@xxxxxxxxx> wrote:
Hi
Sorry for the late reply.
I did take snaps of gv$sesstat for the PQ slave which was doing cell single
block reads but saw nothing strange
The following was sesstat in two snapshots with 60 seconds interval, only
statistics where value - prev_value > 0 is considered
SNAP_DATE NAME
VALUE-PREV_VALUE INST_ID SID
---------------
----------------------------------------------------------------
---------------- ---------- ------
20220614 221231 messages sent
2 7 1221
20220614 221231 cluster wait time
110 7 1221
20220614 221231 blocks decrypted
364 7 1221
20220614 221231 gc hash slot removed
374 7 1221
20220614 221231 prefetched blocks aged out before use
461 7 1221
20220614 221231 pinned buffers inspected
536 7 1221
20220614 221231 dirty buffers inspected
2156 7 1221
20220614 221231 shared hash latch upgrades - no wait
4236 7 1221
20220614 221231 user I/O wait time
5439 7 1221
20220614 221231 non-idle wait time
5549 7 1221
20220614 221231 hot buffers moved to head of LRU
6684 7 1221
20220614 221231 calls to kcmgas
40158 7 1221
20220614 221231 free buffer inspected
66597 7 1221
20220614 221231 cell flash cache read hits
66681 7 1221
20220614 221231 physical read requests optimized
66681 7 1221
20220614 221231 physical read total IO requests
66681 7 1221
20220614 221231 physical reads
66681 7 1221
20220614 221231 physical read IO requests
66681 7 1221
20220614 221231 gc local grants
66681 7 1221
20220614 221231 physical reads cache
66681 7 1221
20220614 221231 free buffer requested
66681 7 1221
20220614 221231 gc remote disk read
66681 7 1221
20220614 221231 consistent gets from cache
68471 7 1221
20220614 221231 consistent gets examination
68471 7 1221
20220614 221231 consistent gets
68471 7 1221
20220614 221231 consistent changes
68471 7 1221
20220614 221231 transaction tables consistent reads - undo records applied
68471 7 1221
20220614 221231 session logical reads
68471 7 1221
20220614 221231 non-idle wait count
161928 7 1221
20220614 221231 file io wait time
54386966 7 1221
20220614 221231 physical read bytes
546250752 7 1221
20220614 221231 cell physical IO interconnect bytes
546250752 7 1221
20220614 221231 physical read total bytes
546250752 7 1221
20220614 221231 physical read total bytes optimized
546250752 7 1221
20220614 221231 logical read bytes from cache
560914432 7 1221
35 rows selected.
Thanks
On Wed, Jun 15, 2022 at 10:10 PM Tanel Poder <tanel@xxxxxxxxxxxxxx> wrote:
Check the v$sesstat metrics, there's nowadays plenty of feedback from storage
cells telling how "deep" the offload processing actually was or if there was
pushback/passthrough caused
Common causes for smart scan "hiccups" with single block reads:
1. chained rows
2. migrated rows for some operations (update, maybe select for update etc)
3. uncommitted transactions in blocks (or not trivially cleanable blocks
in storage)
But v$sesstat metrics will tell you more.
Here's an article from 12 years ago, where chained rows caused problems for a
smart scan.
*
http://tech.e2sn.com/oracle/exadata/performance-troubleshooting/exadata-smart-scan-performance
Things have somewhat improved by now, but back then the "cell chained rows%"
metrics didn't even exist, had to reason by looking at other metrics, like the
"table fetch continued row" and the difference of "processing depth" at cache &
txn layer vs data layer in the storage cells:
869, DWH , STAT, table fetch continued row ,
3660, 732,
869, DWH , STAT, cell blocks processed by cache layer ,
5428, 1.09k,
869, DWH , STAT, cell blocks processed by txn layer ,
5428, 1.09k,
869, DWH , STAT, cell blocks processed by data layer ,
3625, 725,
--
Tanel Poder
https://learn.tanelpoder.com
On Tue, Jun 14, 2022 at 6:26 PM Ls Cheng <exriscer@xxxxxxxxx> wrote:
Hi all
I have a strange situation where a very simple query (no join, a single FROM
table) sometimes is fast (seconds) and sometimes slow (hours). After digging a
bit It seems that one of PQ slave instead of accessing the table using cell
smart table scan is accessing by cell single blocks.
This is 19.10, has anyone observed or faced such a problem?
Thanks