Mining AWR: Statistics Metrics verses Statistics

Update: second posting on wait event metrics at:

In this first installment on metric tables we’ll look at system statistics. I future posts we’ll look at  waits (events) , file metrics, and possibly some other metrics after that. Here are  the tuning metrics tables (SQL  stats are not in “metric” tables per say)

(*DBA_HIST_…_HISTORY views are sort of confusing. AFAI remember they were storing alert history, but apparently they are used for adaptive thresholds – an area for future investigation)

I’ve noticed a number of people posting queries using DBA_HIST_SYSSTAT instead of DBA_HIST_SYSMETRIC_SUMMARY which leads me to believe that there is some confusion or lack of information on the metric tables.

Oracle 10g introduced metric tables which compute deltas and rates of statistics thus hugely simplifying the ability to answer simple questions like “what is the I/O rate on my databases right now.” This question, before 10g, was surprisingly tedious to answer. To answer the question one would have to query v$sysstat  for example:

Select value from v$sysstat where name=’physical reads’;

but querying v$sysstat just once fails to answer the question but instead answers the question “How much I/O has been done since the database was started”. To answer the original question one would have to query v$sysstat twice and take the delta between the two values:

  • Take value at time A
  • Take value at time B
  • Delta = (B-A)
  • and/or get Rate = (B-A)/elapsed time

Getting these deltas and rates could be a pesky task especially working with a customer over the phone. Then 10g Oracle introduced metric tables which answer the questions in one single query using


such as

from v$sysmetric
where metric_name='Physical Reads Per Sec';
---------- ----------------- ------------
654.6736 Reads Per Second          5959
134.9835 Reads Per Second          1515

Notice that the query returns 2 rows. The first row is the the last minute (ie 59.59 seconds) and the second row is the last 15 seconds (ie 15.15 seconds). Oracle collects both the deltas and rates for 60 second and 15 second intervals.

Oracle has the average, maximum, minimum for the values for the last hour in


that one can query like:

where metric_name='Physical Reads Per Sec';

---------- ---------- ---------- ------------------
3.71784232          0 .076930034         .478529283

Also for the last hour Oracle stores the 60 second intervals and for the last 3 minutes the 15 second intervals in



Then for the last week by default, Oracle saves the values for each hour including the maximum, minimum, average, stddev etc in



One issue with using

  • V$SYSMETRIC – last 15 and 60 seconds
  • V$SYSMETRIC_SUMMARY – values  last hour (last snapshot)  like avg, max, min etc
  • V$SYSMETRIC_HISTORY – last hour for 1 minute, last 3 mintes for 15 second deltas
  • DBA_HIST_SYSMETRIC_SUMMARY – hour summaries for last week.

is becoming familiar with the statistics names which are different from v$sysstat. We can look at


For the group_names (statistic definitions)

  • System Metrics Short Duration – 15 second delta stats (41 10gR2, 47 11gR2) – not sure I’d ever use short duration
  • System Metrics Long Duration – 60 second delta stats (135 10gR2, 158 11gR2) – might as well just use long duration

For trending data over multiple days, the view DBA_HIST_SYSMETRIC_SUMMARY can be used.  The view has a history of all the System Metrics Long Duration statistics.  If you store multiple databases in the same AWR repository you can check the statistics available to a particular DBID with the view DBA_HIST_METRIC_NAME. The view DBA_HIST_SYSMETRIC_SUMMARY  can be queried easily for trending metrics, such as the simple query for bytes read by the database per second:

select   dbid, to_char( begin_time ,'YYYY/MM/DD HH24:MI'),
from     dba_hist_sysmetric_summary
where    metric_name= 'Physical Read Total Bytes Per Sec' /* and DBID=[dbid if share repository] */
order by begin_time;

Compare this to the same query on DBA_HIST_SYSSTAT (note there are a lot of stats in v$sysstat)

with stats as (
           select sn.dbid,
                  to_char(cast(begin_interval_time as date ), 'YYYY/MM/DD HH24:MI') btime,
                  Lag (st.value) OVER( PARTITION BY st.stat_name,st.dbid ORDER BY sn.snap_id)
                  st.value     value_end,
                  (cast(end_interval_time as date) - cast(begin_interval_time as date )) * (24*60*60) delta
                  DBA_HIST_SYSSTAT  st,
                  DBA_HIST_SNAPSHOT sn
                    sn.snap_id=st.snap_id and
                    sn.dbid=st.dbid and
                    (st.stat_name= 'physical read total bytes')
           order by begin_interval_time
          dbid, btime,
          round((value_end-value_beg)/delta) rate_per_sec
   from stats
   where (value_end-value_beg) > 0

Its a bit disconcerting to note that the above two queries don’t return the exact same data on my laptop. If it was roughly the same that would be fine, and in general the stats are the similar but there are cases where they differ dramatically.  I don’t see anything obvious in the way the queries are written.  Possibly has to do with database bounces or the way the database is affected by the laptop’s sleep and hibernate modes. Will have to look into this farther.

One trick to make the data easy to load into Excel is to use the html output format and spool to a file with an “.html” extension

SET markup HTML on
spool output.html

List of DBA_HIST views


  1. Trackbacks

  2. July 8th, 2011: Log Buffer #228, A Carnival of the Vanities for DBAs | The Pythian Blog

  2. Paresh
    July 10th, 2011 at 05:21 | #1


    Thanks for the post. One reason for not using could be resistance to change!! It is just easier to calculate the delta and divide by time rather than remembering all the variations (60 sec, 15 sec, hourly) etc. across different views.


    November 5th, 2011 at 18:06 | #2

    Hi Kyle

    I am also surprised why many people still uses sysstat instead of metrics to obtain rate of X per second, but as you I have encounter quite a few situations where both value differs so I dont really know which of these view returns more accurate value.

    Have you found out why they are very different sometimes? I have observed same behaviour in 10g and 11g.


  4. November 5th, 2011 at 21:31 | #3

    I don’t know why the stats sometimes differ.
    There have been a number of posts on Oracle-L over the years as well as some blog posts. I haven’t looked into it closely. My guess is that the metrics are better. The metrics are getting more road testing as the are used by the alerting system in Oracle.

You must be logged in to post a comment.