in Oracle

StatsPack and AWR Reports — Bits and Pieces — 3

This is the third post in this series.
Post 1 is here
Post 2 is here

Note : Some figures / details may be slightly changed / masked to hide the real source.

Identifying Unusual Events / Behaviours / Applications

Here I find a few “unusual” events and mark them out.

Extract A : 11.2 AWR

Snap Id Snap Time Sessions Cursors/Session
Begin Snap: …. 03:00:57 107 .7
End Snap: …. 07:00:07 114 .9
Elapsed: 239.17 (mins)
DB Time: 22.61 (mins)

Top 5 Timed Foreground Events

    Event Waits Time(s) Avg wait (ms) % DB time Wait Class
    DB CPU 1,332 98.16
    SQL*Net more data to client 49,701 20 0 1.50 Network
    SQL*Net more data from client 213,915 5 0 0.34 Network
    db file scattered read 1,159 1 1 0.08 User I/O
    db file sequential read 7,547 1 0 0.07 User I/O

    The two “SQL*Net more data” sets of waits are the unusual events.
    The Time on SQL*Net more data to/from client is negligible isn’t it ? So, should I be concerned ? Over a 4 hour period, only 20seconds were on “SQL*Net more data to client”. Time on “SQL*Net more data from client” is much lower at a total time of 5seconds only. So “Time based” tuning would ignore these two waits.

    Foreground Wait Events

    Event Waits %Time -outs Total Wait Time (s) Avg wait (ms) Waits /txn % DB time
    SQL*Net more data to client 49,701 0 20 0 37.78 1.50
    SQL*Net more data from client 213,915 0 5 0 162.62 0.34
    db file scattered read 1,158 0 1 1 0.88 0.08
    db file sequential read 7,550 0 1 0 5.74 0.07
    SQL*Net message to client 652,102 0 0 0 495.89 0.04
    SQL*Net message from client 652,102 0 183,327 281 495.89

    Not that Oracle treats “SQL*Net message from client” as an idle wait so the 183,327seconds of wait time do NOT appear in the Top 5 Timed Foreground Events list.

    I would draw attention to the high number of “more data from client” waits and the correlation with the “message from client” waits. Either extremely large SQL statements or PLSQL blocks are being submitted very frequently or row inserts with very large array sizes (number of rows per insert) are being received. In this case, further investigation reveals an ETL loader that does bulk inserts of a number of rows per array. If we need tuning, tuning the SDU may help.
    Similarly the “more data to client” indicates large data sets are being returned. The numbers of columns and rows per every “send” are high.

    Instance Activity Stats (from the same AWR)

    Statistic Total per Second per Trans
    Requests to/from client 654,738 45.63 497.90
    SQL*Net roundtrips to/from client 654,740 45.63 497.91
    bytes received via SQL*Net from client 1,793,072,463 124,950.54 1,363,553.20
    bytes sent via SQL*Net to client 552,048,247 38,469.57 419,808.55
    logical read bytes from cache 762,514,227,200 53,135,924.61 579,858,727.91
    physical read total bytes 8,772,479,488 611,311.62 6,671,087.06
    physical write total bytes 25,334,243,328 1,765,420.76 19,265,584.28
    redo size 6,373,204,848 444,117.79 4,846,543.61

    1.793billion bytes received in 654K SQL*Net trips is 2741bytes per trip received at 45 messages per second. Given that it is still only 2,741bytes per trip, possibly the array size could also be tuned with the SDU and TDU.

    So, this is an AWR that doesn’t call for tuning but reveals information about how the database is being used. Large number of rows (large number of columns) being inserted and retrieved in each call. The Performance Analyst needs to be aware of the nature of the “application” — here it is not OLTP users but an ETL job that is the “application”. Although the database had more than a hundred sessions a very small number of sessions (possibly 4 ?) were active doing ETL and checking the status of ETL tables during this window.

    Would you care to analyze the other statistics I’ve listed — bytes read and bytes written ?

    Extract B : 10.2 AWR
    Snap Id Snap Time Sessions Cursors/Session
    Begin Snap: 00:00:07 228 4.5
    End Snap: 23:00:06 232 4.3
    Elapsed: 1,379.97 (mins)
    DB Time: 11,543.24 (mins)
    I know. It is really bad and most likely meaningless to get an AWR for a 24hour range. (I’ll not go into the details about why the AWR is for 24hours — save that I did NOT ask for a 24hour AWR report).
    Top 5 Timed Events

    Event Waits Time(s) Avg Wait(ms) % Total Call Time Wait Class
    CPU time 258,101 37.3
    db file sequential read 62,150,655 208,148 3 30.1 User I/O
    db file scattered read 28,242,529 141,638 5 20.4 User I/O
    RMAN backup & recovery I/O 1,597,421 37,137 23 5.4 System I/O
    enq: TX – row lock contention 22,276 34,942 1,569 5.0 Application
    The RMAN Backup load is expected. (Why ? Given a 24hour report, I expect RMAN to have run at least once during the day).
    For performance tuning, I would look at the “db file … read” events and identify and analyze SQL statements and the schema.
    What is the “unusual” event here ? It is the “enq: TX – row lock contention“. Over a period of 1,380minutes, there were 22,276 Row-Lock Contention waits. Actually, this application does not have the same load throughout the 23hours. Most likely, it had load for 15hours only. So, we had 22,276 Row-Lock Contention waits over 15hours. That translates to 1,485 waits per hour or one Row-Lock Contention wait every 2.5seconds. Now, that is a very high frequency. Either users are locking each other out for a very short while (1.569seconds per wait on average) or there is/are one or more jobs that run at a very high frequency and update a common “reference” table. I won’t reveal my findings here but analysis of the SQL statements indicates what the “problem” is.
    Now, should “Time based performance tuning” be worried about the 5% of time lost on these waits ? Probably not. But they do indicate something peculiar in the design of this system. There are less than 250 user sessions in this OLTP database but there is/are one or more jobs that is/are locking itself every 2.5seconds — so there is some point of serialisation occurring. Is that job also accounting for CPU time or ‘db file read’ time ? That needs further analysis.
    Both these cases show how a Performance Analyst needs to know how the database is being used. What sort of jobs are submitted, besides OLTP users ?