Case Study Abstract
This case study demonstrates a top-down tuning methodology that was applied to a performance problem involving high CPU usage on a database server. The diagnostics that were collected and analyzed for this problem include: operating system (OS) data such as vmstat and top; statspack, and trace event 10046 with TKProf output. The conclusion from analyzing the collected data was that a specific query was consuming most of the CPU and needed to be tuned. We present various ways of tuning the query as well as verification that the recommended solution solved the problem.
Case History
A large university reported severe performance problems with their email application and other applications that relied upon an Oracle RDBMS server for authentication services. The Oracle database was used solely for user lookups, authentication, and other assorted LDAP queries. It was hosted on a 2-CPU Linux server that also hosted Oracle’s Application Server (iAS) and related daemons.
The email application was installed approximately 6 months prior and ran without incident until some application security patches were installed. The day after the patches were installed, the performance problems started occurring.
The DBAs noticed that this authentication server was often seeing load averages over 20 processes (as shown in uptime or top). Normally, prior to the patch, the load averages were seldom above 1 or 2 processes. The performance seemed to be much worse when “mass mailing” occurred. These mass-mailing jobs sent out email to thousands of recipients and relied upon the authentication/lookup services provided by the database.
The main symptoms reported by users when there was a performance problem were that they were unable to log-in to their email accounts and they experienced timeouts in the application (visible as browser timeouts or client error messages, depending on the mail client used).
The database server version was 9.0.1.5; it had been upgraded from 9.0.1.4 during the same outage period when the application security patches were installed.
Analysis
Summary
Our strategy in diagnosing this problem was to take a “top-down” approach and follow the trail of highest CPU consumption down to the offending processes, and eventually the offending sessions and query. The top-down approach can be summarized as follows1:
1. Verify CPU Consumption on the server
• Verify the observation of high CPU consumption on the affected server and collect top and vmstat data for a period of time (from 30 minutes up to a few hours)
• Find which processes are using the most CPU
• If the top CPU consumers are Oracle processes, then find out which database sessions correspond to these processes2
2. Collect Extended SQL Trace Data (Event 10046)
• Gather SID, SERIAL#, and other information from V$SESSION about these sessions
• Trace the sessions using an extended SQL trace via event 10046, level 12
• Produce a TKProf report and look for the queries having the longest elapsed time
3. Build a Test Case Script for the Problem Query
• Extract the high-impact queries with their bind values and create a test script that will allow us to benchmark changes to the query and verify our progress in tuning it.
• Run the test script on the production system (with the 10046 event set) to validate that it reflects what was seen in step two.
4. Obtain Statspack Snapshots
• Collect statspack snapshots during the same time period as the OS were collected and look for correlations with the session-level data collected in step two. This could easily be done before steps 1 through 3 as well.
5. Tune the Query
• Tune the query and verify performance improvements using the test script
• Implement the changes in production and verify that the solution worked.
• Verify the performance has improved and meets the customer’s requirements
In steps 1 and 2 above, one can see that the overall effort is one that focuses attention down to a particular session, with the ultimate goal of collecting the extended SQL trace information. We are interested in the extended SQL trace information because it will report CPU times, wait times, elapsed times, and bind values3. By having the CPU, wait, and elapsed times, we can account for how the time was spent during each query’s execution. The bind values allow us to put together a reproducible test case using actual values. The bind values will be extremely useful later when we tune the query.
Detailed Analysis
1. Verify CPU Consumption on the server
Before we can start looking at why this problem is occurring we must first understand the problem and verify what the customer is reporting (its symptoms). This can be done by collecting OS statistics during a recent spike in activity due to a mass mailing. The data was collected using Oracle Support’s OSWatcher tool4.
The vmstat output during this time looked like this:
Each line represents a sample collected at 1-minute intervals.
Notice the how the run queues5 rapidly increase from 3 to 25 as the CPU utilization increases from about 50 percent to 100 percent. This shows the importance of knowing the length of the run queue as well as the CPU utilization. When the CPU is pegged at 100% utilization, the severity of the CPU starvation won’t be reflected in the percentage of CPU utilization (its pegged at 100%), but the run queue will clearly show the impact. Similarly, knowing only the run queue will not provide you with knowledge on the exact CPU usage, nor the spread of time usage across system and user modes.
CPU utilization is classified into three types: system (sy), user (us), and idle (id). System mode CPU utilization occurs whenever a process requires resources from the system; e.g.,
I/O or memory allocation calls. In typical OLTP systems, the percentage of system mode CPU utilization is often less than 10 percent; in data warehouses more I/O calls are performed and a higher percentage of system mode CPU utilization is common. User mode CPU utilization accounts for the rest of the time the CPU is busy and not running in system mode. Idle CPU is essentially the time that the CPU is not busy and waiting for work to do.
In this case, we can quickly see that the percentage of CPU utilization in system mode is usually less than 10 percent. This indicates that we are probably not suffering from memory shortages or doing excessive I/Os6. We can further dismiss memory shortages by looking at the paging and swapping statistics in vmstat (only swap-ins (si) and swap-outs (so) are shown here). If no pages are being swapped out during peak time, then there can’t possibly be a memory shortage.
If we were seeing system-mode CPU utilization higher than 15 percent and we weren’t seeing any memory problems, we might begin to suspect excessive I/O calls, very frequent database connections/disconnections, or some anomaly (we would keep this in mind as we collect more data for individual processes; e.g., using the truss command or while looking at RDBMS metrics like logons cumulative per second).
The fact that most CPU utilization is occurring in user mode tells us that whatever is occurring, it’s happening within the application that the process is running, not as part of a system call or action. The question is now, which processes are using the CPU?
To answer this question, we can look at the output of the top command. This command is useful because it shows at a glance the overall CPU statistics and the processes consuming the most CPU (ranked in descending order of CPU usage). The following shows a typical output of the top command during the performance problem:
Notice the load average for the last one-minute has been around 22 (load averages are for the last 1, 5, and 15 minutes). The load averages are an average7 of the run queue and basically means that around 22 processes were ready to run; however, the machine has only four virtual CPUs (two real Intel Xeon Hyperthreaded CPUs). The lines beginning with “CPUx states” show that Linux believes there are 4 CPUs. This means that on average, each CPU has one process running and over 5 waiting to run (22 / 4).
It’s clear from the output of the top command that Oracle database processes are using nearly all of the CPU. We can further clarify the names of these processes by looking at the output of the ps command (or use top’s “c” command in top’s interactive mode):
The ps command8 shows the accumulated CPU time for each process as well as the full name for the process associated with the process ID (PID). The process IDs are important for moving along to the next stage of the diagnostic process where we examine and trace the database sessions that are associated with these PIDs. The full name tells us what kind of process it is. In this case they were all Oracle database processes.
At this point we have verified the following:
• The system is under extreme CPU starvation and very high run queues have formed
• Oracle database processes are responsible for the high CPU consumption
If we had found non-Oracle processes using most of the CPU, then we would need to drill-down into those processes and investigate the reason for such high CPU consumption. Since Oracle processes were clearly the culprits here, we drilled down into what Oracle was doing.
Database Drilldown
Once we’ve verified the problem exists and involves Oracle processes, it was time to determine the cause of the CPU consumption. From the top and ps output, we had a list of PIDs that were responsible for most of the CPU consumption. The next step was to find out more about the sessions in the database that were associated with these processes. We wanted to find out the following:
• Whether any of these sessions were “runaways” (i.e., stuck in an infinite loop and continuously using CPU)
• Which SQL hash values were most commonly seen when the session was active (same SQL statement)
• SID and serial# for tracing purposes
Here is the query that was used to find out more about these sessions:
SELECT /*+ ordered */ p.spid, s.sid, s.serial#, s.userna
me, TO_CHAR(s.logon_time, 'mm-dd-yyyy hh24:mi') logon_time, s.last_call_et, st.value, s.sql_hash_value, s.sql_address, sq.sql_text
FROM v$statname sn, v$sesstat st, v$process p, v$session s, v$sql sq
WHERE s.paddr=p.addr
AND s.sql_hash_value = sq.hash_value and s.sql_Address = sq.address
AND s.sid = st.sid
AND st.STATISTIC# = sn.statistic#
AND sn.NAME = 'CPU used by this session'
AND p.spid = &osPID -- parameter to restrict for a specific PID
AND s.status = ‘ACTIVE’
ORDER BY st.value desc
We noticed the following when we ran this query:
• None of the top CPU consuming sessions had last_call_et greater than 3 seconds any time we ran the query. No runaways…whatever is going on is completing within 3 seconds.
• The same 10 or 12 sessions were at the top accruing CPU time.
• We very frequently saw the same sql_hash_value and SQL statement while the sessions were active
• The statistic CPU used by this session statistic was accruing time steadily
After running this query, we had a list of 12 candidate sessions that we could trace using the extended SQL trace event. We chose the session with the highest accumulated CPU time (CPU used by this session), any one of them would have been a good candidate since their total CPU consumption differed only by a few percentage points.
2. Collect Extended SQL Trace Data (Event 10046)
At this point in the process, we were ready to gather the extended SQL traces and find out exactly what the sessions were doing to use all of this CPU.
The extended SQL trace will help us account for the time spent executing SQL for a particular session. We will be able to account for CPU and wait time as well as see the sequence in which SQL statements are emitted by the application and which of them is taking the most CPU and/or elapsed time. In addition, we’ll obtain the bind values associated with a query so that we can use them to create our test script. We have to ensure the statement uses binds because the CBO might choose different execution plans if we use literals instead.
With the session ID and serial number of the session obtained from the V$SESSION queries we ran earlier, we started the 10046 trace as follows:
SQL> SELECT p.spid, s.sid, s.serial# FROM v$session s, v$process p
WHERE s.paddr = p.addr AND p.spid = 24078
SPID SID SERIAL#
------------ ---------- ----------
24078 18 5
SQL> begin dbms_system.set_ev(18,5, 10046,12,''); end; -- trace on
-- collect trace information for approximately 15 minutes during the problem
SQL> begin dbms_system.set_ev(18, 5, 10046,0,''); end; -- trace off
This produced a trace file in the user_dump_dest directory. We used the tkprof 9utility to format the trace file and sort the SQL statements in descending order of highest fetch elapsed time as follows:
$ tkprof ora_24078.trc 24078.prf sort=fchela
$ cat 24078.prf
<some lines removed for brevity>
SELECT /*+ USE_NL(store) USE_NL(dn) INDEX(store EI_ATTRSTORE) INDEX(dn EP_DN)
ORDERED */ dn.entryid,store.attrname, NVL(store.attrval,' '),
NVL(store.attrstype, ' ')
FROM
ct_dn dn, ds_attrstore store WHERE dn.entryid IN ( (SELECT /*+ INDEX( at1
VA_objectclass ) */ at1.entryid FROM CT_objectclass at1 WHERE at1.attrvalue
= :0) UNION (( (SELECT /*+ INDEX( at1 VA_mail ) INDEX( at2 VA_objectclass
) */ at1.entryid FROM CT_mail at1, CT_objectclass at2 WHERE at1.attrvalue
= :1 AND at2.attrvalue = :2 AND at2.entryid = at1.entryid) UNION (SELECT
/*+ INDEX( at1 VA_mail ) INDEX( at2 VA_objectclass ) */ at1.entryid FROM
CT_mail at1, CT_objectclass at2 WHERE at1.attrvalue = :3 AND
at2.attrvalue = :4 AND at2.entryid = at1.entryid) )) ) AND ( (dn.parentdn
like :bdn ESCAPE '\' OR (dn.rdn = :rdn AND dn.parentdn = :pdn)) ) AND
dn.entryid = store.entryid AND dn.entryid >= :entryThreshold AND
store.attrkind = 'u' AND store.attrname NOT IN ('member', 'uniquemember')
call count cpu elapsed disk query current rows
------- ------ -------- ---------- ---------- ---------- ---------- ----------
Parse 555 0.09 0.83 0 0 0 0
Execute 555 0.42 0.78 0 0 0 0
Fetch 555 114.04 385.03 513 1448514 0 11724
------- ------ -------- ---------- ---------- ---------- ---------- ----------
total 1665 114.55 386.65 513 1448514 0 11724
Elapsed times include waiting on following events:
Event waited on Times Max. Wait Total Waited
---------------------------------------- Waited ---------- ------------
SQL*Net message from client 556 1.94 134.68
SQL*Net message to client 555 0.00 0.00
db file sequential read 513 0.65 6.68
SQL*Net more data to client 169 0.00 0.02
latch free 323 1.03 59.85
The top SQL statement found in the TKProf report (sorted by fetch elapsed time) was the same one we saw frequently while querying V$SESSION (hash value of 1656209319).
Let’s analyze the TKProf to see how the time was spent for this SQL statement:
1.) An average of 11,724 rows / 555 fetch calls = 21 rows/call is returned
2.) An average of 1448514 buffer gets /555 executions = 2,609 buffer gets/execution
3.) The elapsed time for this statement was 386.65 seconds
4.) The CPU time for this statement was 114.55 seconds
5.) The wait time should be:
wait time = elapsed time – CPU time
wait time = 386.65 – 114.55 = 272.10 seconds
6.) The sum of the wait times shown in the area under “Elapsed times include waiting on the following events” is 66.55 seconds.
We exclude the time listed as “SQL*Net message from client” because this wait time is not part of the elapsed time for the query. When we examined the trace file, we saw that this was the database “idle” time in between the completion of this query while waiting for the next query to be processed. In effect, this is the client and network time (a.k.a. “think time” from the database’s point of view)10.
In this case we ignore the time in between database calls (“SQL*Net message from client”, 134.68 seconds) because we know the system is starving for CPU. If we hadn’t seen any bottlenecks at the OS or database stacks, we would pay attention to this inter-call time because it could strongly indicate a bottleneck in the middle tier (not a database performance problem). The CPU saturation is the main bottleneck – clearly, the machine is not anywhere near idle that we need to look elsewhere for a performance problem!
It’s interesting that the wait time didn’t add up to the 272.10 seconds we anticipated…where did the “missing” time go? Because this system was undergoing severe CPU starvation and the run queues were very high (over 20), processes were spending a considerable amount of time waiting in the run queue. While they were in the run queue, Oracle didn’t accrue any CPU time, nor was it accruing any wait time from the database point of view (i.e., the session wasn’t waiting for “db file sequential read” or anything else because it wasn’t sleeping). It’s as if time stood still for the session while it was on the run queue, but, in fact time was passing (“elapsed time” or “wall clock” time) – it just wasn’t measurable by Oracle until the call ended11.
What about the “latch free” wait event – was that important? Normally, one would say ‘yes’ and would focus on why we’re seeing latching problems. In cases where the system is facing CPU starvation, apparent latch contention can flare up as a byproduct of the CPU starvation12.
It’s also possible that we have real latch waits on the cache buffers chains latch due to block contention (“hot blocks”). Considering that this query executed 555 times in 17 minutes by a dozen or so concurrent sessions, it’s quite possible that they were contending for the same index root and branch blocks (and hence the cache buffers chains latch that protects the buffer headers for those blocks).
However, this points to the same root cause as the high CPU: the query needs to be tuned.
Examining the Query’s Execution Plan
The runtime execution plan is shown just below the statement statistics in TKProf and looks like this:
The areas of interest in this plan are:
• the top line that shows 19 rows were returned
• the line in the middle: “64436 INDEX RANGE SCAN (object id 55349)”
This means that an index range scan occurred and this scan visited 64,436 rows but ultimately only 19 rows were returned by the query. A lot of work was done for very few rows. Typically, OLTP systems use indexes to minimize the amount of I/Os required to fetch the necessary rows. We can be sure that the average of 2,600 or so buffer gets per execution occurred due to this index. Buffer Gets is a measure of how much work the database is performing for each SQL statement and could also be referred to as Logical I/Os (LIOs). This is a strong indication that we should attempt to tune this query and reduce the number LIOs and thereby reduce the CPU consumption (and, as a bonus relieve the latch contention).
Our next step will be to put together a simple test script so we can begin tuning this query.
3. Build a Test Case Script for the Problem Query
In order to build an accurate test case script for this query, we will need to set the bind values to some typical values that were captured in the extended SQL trace. The bind values are found in the trace file as such:
When we extract the bind values and set up the variables, we obtain the following script that can be run in SQL*Plus:
Our test case script produced the following trace:
We can now compare the results of the trace generated while running our test case script to the original trace we obtained of an actual running process. We find that they have:
• The exact execution plan
• The VA_OBJECTCLASS index is traversing approximately the same number of rows (64,000 rows)
• The buffer gets per execution is approximately the same at 2,600
This tells us the test script is representative of the actual query and we can confidently use it for tuning purposes13.
4. Obtain Statspack Snapshots
Before we begin tuning the query using our test script, it will be useful to briefly consider overall database performance statistics collected using Statspack.
Statspack snapshots were taken of the database to get an overall picture of database performance at time of the problem (while collecting OS statistics and 10046 tracing).
Ideally, we would compare the statspack reports when performance was good to the ones during the performance problem. Doing this kind of comparison is extremely valuable because we would be able to see a change in the CPU, wait event profile, and SQL statement performance and then pursue the reasons for the specific changes. Unfortunately, in this case statspack wasn’t installed until after the performance problem was encountered, so we have no performance data from before the patch was applied.
It could be argued that this step was rather unnecessary since we already knew that CPU was being used by a number of sessions and the next step would have been to simply trace those sessions and see what they were doing. While this is a valid point, obtaining statspack snapshots is easy and will give us some overall context as well as provide a good sanity check that what we see for the overall database is present in an individual session’s SQL trace file. Capturing this data also allows us to quantify the effect of this single statement’s resource usage against that of the entire instance.
Accounting for Database Time
During the performance problem we obtained the following excerpts from the statspack report’s header section:
It’s clear that the top wait event is related to latch acquisition. This may tempt some tuners to try and resolve a latching problem and find out which latch is the main point of contention. However, wait time is really just part of the total time that must be accounted for in the database. The total active14 time spent in the database (a.k.a DB Time in later versions) by foreground sessions15 is the following:
Total DB FG Time = Total DB FG CPU time + Total DB FG Wait Time
Unfortunately, since this is a 9.0.1.x database, Statspack did not incorporate CPU time into the “Top 5” list. To obtain the CPU for foreground Oracle sessions, you will need to look in the “Instance Activity Stats” section of the statspack report for the value of CPU used by this session16 as shown below.
Now, this clearly reflects the fact that the database is spending most of its time on the CPU (or waiting for CPU). Some questions we might have at this point are:
1.) What is using this CPU?
The first place to look is the section “SQL ordered by Gets”. Looking here we see the statement we identified using the 10046 trace accounts for 90% of the CPU usage on this system in this snapshot period. This confirms this is the statement to tune (more detail on this below).
2.) Why did we have 5,291 seconds of database time when the elapsed time for this report was just 1,576.8 seconds (26.28 min)?
The total DB time depends on the number of sessions waiting in the database (which is only limited by how many sessions could possibly connect to the database) and the number of CPUs available on the machine. As an extreme example, one could have 1,000 sessions all waiting for an enqueue for an entire hour and the total database time would be:
1,000 sessions X 3,600 seconds of wait /session = 3,600,000 seconds
of DB time
Or, one could have no sessions waiting but four CPUs and four sessions busy using the CPUs for one entire hour:
4 sessions X 3600 sec of CPU / session = 14,400 seconds of DB time.
In this report we can see that there were approximately 200 sessions17 during the report period (see the report header section above) and a total DB time possible
of:
200 sessions X 1,576.8 seconds / session = 315,360 seconds of DB possible.
The actual DB time is much lower because the majority of sessions were idle (there were hundreds of idle sessions and just about a dozen busy ones).
3.) How does the 3,319 seconds of CPU time compare to the total amount of CPU time that was available?
To answer this one has to work out what the total possible CPU time was for the duration of the statspack report and then compare that to the total amount of CPU utilized. For this system we were investigating, there were two hyperthreaded CPUs which gave us a total of 4 CPUs for the statspack report duration of 26.28 minutes, for a total possible amount of CPU of:
4 X 1576.8 seconds = 6307.2 seconds of CPU available
The total CPU utilization by the database was then: 3,319 / 6307.2 = 53%.
Although this is a large value, it appears to contradict our view from top and vmstat that the database was using close to 100% of CPU. Unfortunately, we didn’t find the reason for this discrepancy but it may be related to the CPUs being “Hyperthreaded” rather than actual independent CPUs. We believe the database’s CPU metrics may have somehow been underestimated by a factor of two.
4.) Why are latch waits so prominent?
When CPU is scarce, it is quite likely that a session will be holding a resource such as a latch or enqueue when it is taken off the CPU and put on the run queue. This means that another session that needs that resource will have to wait for it much longer than it otherwise would. In fact, it has to wait until the holder is back on the run queue long enough to finish what it was doing and release the latch or enqueue. This point highlights the importance of understanding the overall performance picture from an OS standpoint before jumping to conclusions in statspack or a 10046 trace. When CPU starvation is involved, one should usually avoid putting too much credit in the wait events and instead seek to understand what is consuming the CPU18. Of course, as was discussed before, hot blocks due to the large number of concurrent buffer gets from the top query could cause the latch contention, but at this point they shouldn’t be the main focus of our effort.
Finding the Top SQL statements in Statspack
Although we have already identified a particular SQL statement as a candidate for tuning, we will briefly look at the top SQL statements reported by Statspack to see if they agree with what we saw for an individual session. We’ll be sure that we’re on the right track if the top statement in the trace agrees with what we see as the top statement for the overall instance in Statspack
We’ll look at the statspack report’s section called “SQL ordered by Gets” to get an idea of the most expensive SQL statements from a CPU point of view. Buffer Gets is a measure of how much work the database is performing for each SQL statement and could also be referred to as Logical I/Os (LIOs). Each buffer get may or may not result in a physical read from disk, but will definitely incur CPU time as well as possibly experiencing latch or buffer pin waits (a.k.a. buffer busy waits). Here is an excerpt from the Statspack report:
Some key things we notice here:
1.) The statement with hash value 1656209319 was responsible for 74% of the total buffer gets
2.) This statement was responsible for 3,007 / 3,319 = 90% of the CPU consumption in the database.
3.) Each execution used: 3,007 cpu sec/15,642 exec = 192 msec CPU/execution
4.) The statement was doing approximately 2,642 buffer gets for each execution.
5.) The second and third statements were not even close to the first one in terms of total CPU time consumed.
6.) The statement’s total elapsed time was 8,972.36. Huh? How could we wait longer than the DB Time of 5,291 seconds?
The reason is that while a session is executing this statement, it could be put back and forth on the run queue as the OS schedules other processes to run. Each time the process goes on the run queue, the elapsed time continues accruing, but not CPU time. Oracle has no idea it was on the run queue, but eventually when the call finishes, the elapsed time reflects the total execution time including the time on the run queue.
Unfortunately, the statspack report truncates the SQL text and there were other statements that looked very similar to this. So, we had to verify that this is really the same statement that we saw earlier in the TKProf output19. We did this by looking in the raw 10046 trace file we collected and ensured that the hash value in the statspack report was the same one we saw in the 10046 trace:
Then, we compared the full SQL text in the raw trace with the one in the TKProf. We verified this was a match, so we were sure that what we saw in TKProf is the same high-CPU statement that was seen in the statspack report.
At this point we saw that our V$SESSION queries, 10046 trace / TKProf, and Statspack all showed this query as being the CPU consumer. This agreement between these independent data collections gave us confidence that this single SQL statement was the root cause for the performance problems in the database.
Having a particular SQL statement in mind as being the cause for the CPU consumption, we might ask ourselves, “Why did this query cause problems after the patch was applied?” We can imagine other related questions:
1.) Was this query present before the patch was applied?
We don’t know since there are no statspack snapshots prior to the patch
2.) Was this query much more efficient (much less CPU / execution) prior to the patch; did its execution plan change?
Again, we don’t know if it was even present, but considering the hints on the statement (index, join type, and join order), its extremely unlikely that the CBO would choose a different plan.
3.) The instance needs to be tuned using the init.ora parameters
This very seldom yields significant results in cases like this, but we can quickly review the statspack report’s init.ora section for glaring problems.
4.) The query was present before the patch and some strange Oracle bug is involved that is making this query much more expensive
This is possible, but unlikely given the database version changed very little (from 9.0.1.4 to 9.0.1.5). To determine this, we’ll have to install 9.0.1.4 and 9.0.1.5 databases on a test system and run the same test query to compare.
5.) The query is due to the patch and needs to be tuned despite the efforts of the developer that hinted this query.
Quite possibly true and really the quickest way forward at this point.
Which scenario is right? We won’t know unless we talk to the application developers and find out if this query is new or its execution plan is expected. In any case, we will keep our focus on the SQL statement since it clearly is responsible for most of the CPU time in the database.
Reviewing the Init.ora settings using Statspack
Finally, before we leave our discussion about Statspack, it is worth considering the init.ora settings shown in the last section of the statspack report to see if any unusual or undocumented settings are found that could affect query performance or CPU utilization. Here is a list of some noteworthy ones in this report:
The parameters that may affect performance are:
_optim_peek_user_binds = false
This setting determines whether the optimizer will “peek” at bind values when a statement is parsed. Potentially better plans could result with this set to true. No one could recall why this was set. The application developer did not recommend it. However, it’s unlikely this would solve the immediate problem with the identified query since that query is thoroughly hinted and the CBO wouldn’t choose a different plan even if it knew the literal values.
pga_aggregate_target = 33554432
This parameter could affect the way the optimizer costs certain operations such as hash joins and influence the CBO’s execution plans. It could also affect the efficiency of those operations at runtime depending on whether there was sufficient memory
available. If insufficient memory is allocated, then Oracle may have to use temporary segments to complete certain operations (inefficiently). To determine this, one should check the following statistics in the report:
Ideally, all operations are performed in memory – shown in the statistics for “workarea executions –optimal”. Onepass and multipass operations indicate some use of temporary segments; some small amount of onepass operations is tolerable, but multipass operations are to be avoided if possible.20
timed_statistics = true
This indicates that operations are being timed in the database; we knew this already but it’s worth mentioning that without this, none of our response time analysis could be accomplished.
workarea_size_policy = auto
This is needed to make the pga_aggregate_target setting go into effect.
Some settings are known to affect CPU utilization when they are misused or are defective. We did not find any in this case, but here is what we looked for:
_spin_count = <some number larger than default of 2000>
This sets the number of times Oracle will try to acquire a latch before going to sleep for a period of time in the hope that when it wakes up, the latch will be free. If latches are held for a short period of time, this optimizes performance by avoiding a premature sleep. High CPU consumption can result if those seeking silver bullets for a latch contention problem set this value too high.
Our review of init.ora parameters doesn’t reveal anything that would explain the high CPU consumption. It looks like we’ll have to tune this query to use less CPU or find out if the application could execute it less frequently.
5. Tune the Query
To Tune or Not To Tune
At this stage we have identified a specific SQL statement as the cause of the CPU consumption and overall performance problems. Our next task is to find ways to reduce this query’s CPU consumption. We have two possible approaches for doing this:
1.) Execute the query less often (maybe the application patch is unintentionally executing the query more often).
2.) Tune the query to make it use less CPU per execution
We analyzed the execution patterns and the value of bind variables in our extended SQL trace and found no evidence of repeating queries for the same values or other behavior that would suggest “over execution” of the query. The application developers looked at application log files and came to the same conclusion: the query is executing the proper number of times for the incoming requests.
This leaves us with the need to tune the query and make each execution more efficient.
Query Tuning Strategy
Query tuning is typically a non-trivial effort, especially when we have a query with various subqueries involved and a schema that is unfamiliar to us. One approach we could take is to recognize that the CBO may actually be able to find a better plan if only it were given a fair chance. We say “fair” because in this query the CBO’s hands were effectively tied: it had no control over the access paths (INDEX hints were used), join types (USE_NL hints), and join order (ORDERED hint). What if the developer simply made a mistake and chose the wrong sets of hints?
At this point we could apply a tuning technique that might be thought of as the “Judo” 21 query tuning technique: we will use the CBO’s effort and strength to tune the query by giving it more information about the query. We will also explore using session parameters to nudge the CBO to alternative paths.
For this technique to work, we want the optimizer to have all the information about the values in the predicate as possible. The more it knows about these values, the more accurate will be its selectivity estimates, and ultimately its estimate of the cost of obtaining rows for each step of the plan. Selectivity calculations are greatly influenced by the statistics collected for the objects in the query, especially the perceived distribution of data. By default, the optimizer assumes a uniform distribution of data within a column. When histograms are present, the optimizer may accurately determine the selectivity of the query despite uneven data distribution.
However, even with proper histograms to correctly estimate the selectivity for individual predicates, the optimizer still makes the assumption that predicates are independent of each other when computing the selectivity of the entire query. This is not always true and can lead to gross errors in the selectivity calculations. To mitigate this effect we can use dynamic sampling to let the CBO apply the predicates in the query and count how many rows are expected from each rowsource. This permits the CBO to obtain very accurate selectivity values and leads to an execution plan that has realistic cardinality estimates22 and an efficient plan.
In addition to using dynamic sampling, one can also attempt to get a better plan by using parameters such as FIRST_ROWS_K or OPTIMIZER_INDEX_COST_ADJ to bias the optimizer. These should be used only at the SESSION level to give the CBO different paths to cost that it might have otherwise avoided. Once a better plan is found, one can influence the plan using hints or outlines (The use of parameters such as OPTIMIZER_INDEX_COST_ADJ at the instance level are strongly discouraged because it could have detrimental effects on other plans that are performing well and could make future tuning efforts very difficult).
To implement the “Judo” technique, and give the CBO a good fighting chance at finding a better plan, we ensured statistics were collected (large sample size and histograms) and changed the test script to do the following:
• Removed all of the hints from the SQL
• Replaced the bind values with literals23
• Used dynamic sampling
Keep in mind that our intentions were to see if we could obtain a better plan, we weren’t going to require the use of literals or dynamic sampling in production. We simply wanted to know if there was a better plan that was available to us quickly with minimal effort.
Here is the modified test case script:
You might have noticed some unusual parameters in the script. In order to force the use of dynamic sampling in this version of the database, we had to resort to some undocumented parameters:
_dyn_sel_est_on = true: permit dynamic sampling
_dyn_sel_est_num_blocks = 1000: the number of blocks to sample
In later versions, we could enable dynamic sampling using:
optimizer_dynamic_sampling = 10
The hints were easily disabled by replacing the “/*+” with “/*~+” and the literals were inserted in place of the bind values.
After executing the test case and producing another TKProf report, we see the following improved plan:
This plan is significantly better than the previous one. We observe:
• The CBO chose a different index than the hinted plan
• The total gets for this query is 63 compared to about 2,600 for the previous one
• The amount of CPU used per execution is less than 10 milliseconds (below the 10 millisecond resolution in the trace file) compared to approximately 206 milliseconds of CPU per execution
Why did the use of the ST_OBJECTCLASS index improve the performance of this query?
This index included the two columns that were in the predicates of the various subqueries (ENTRYID and ATTRVALUE) instead of having only one column (ATTRVALUE). The ATTRVALUE column used in the VA_OBJECTCLASS index was actually a less selective column (meaning there were few distinct values for this column, 104 out of 829,475 rows actually) so the range scan on that index had to read many index blocks based solely on ATTRVALUE and look for rows matching the value of ENTRYID.
The ENTRYID column had 196,303 distinct values and therefore the ST_OBJECTCLASS index allowed Oracle to navigate directly to a leaf block containing both desired values and scan from there.
We were fortunate that our technique yielded a better execution plan quickly25. Once we had a better plan, this question arose: “what can we do to ensure the application’s queries use this plan?” We have the following potential solutions available to us:
One may wonder if rebuilding the existing VA_OBJECTCLASS index might improve performance. After all, isn’t it possible that this index is fragmented with many deleted rows and Oracle has to read many sparsely populated index leaf blocks just to satisfy the query? It is possible but unlikely because the patch application did not perform any DML on these tables (we learned this from the developer). But, just to confirm, we investigated the index statistics to be sure if this would help or not:
This query shows us that only 2,428 leaf rows are deleted in this index. The “bad” part of our execution plan is returning 64,430 rows, so in the extreme, worst case that all of the deleted rows happened to be in the range returned by our test query, the deleted rows only account for about 4% of those rows. These deleted rows are more likely spread around the entire index and the actual number of deleted rows that were skipped is much smaller than 4%. Simply rebuilding the index wouldn’t have improved the query’s performance (and, it might have wasted time and reduced the customer’s confidence in our ability to solve this problem!).
Instead of just rebuilding the index, we chose to drop and recreate the VA_OBJECTCLASS index to include the ENTRYID and ATTRVALUE columns. We chose this solution because it was easy to do, required no application change (the SQL with the same hints will still work), and presented a low risk of negatively impacting the performance of the application. There was a possibility that this new index would be seen by the CBO as bigger and more expensive to use in some queries, thereby changing some plans, but this was a small risk since most queries were hinted. In any case, this confirms the importance of verifying the performance improvement and detecting any regressions that might have been caused by the solution.
Solution Verification
After dropping and recreating the VA_OBJECTCLASS index on the test instance to include the columns we needed, we re-ran the tests (on the customer’s test system) and verified that the performance was improved as we expected. At the next planned outage, the index was rebuilt in production.
Subsequent analysis from the application development staff found the query had NOT been in use prior to the application patch upgrade, so in effect, this was a new, poorly tuned query.
Conclusion and Learnings
With the index rebuilt, we needed to observe the system to see if this solved the problem. During a time of peak load with mass mailings occurring, the customer said that the run queues on the system became less than 1. The system was now running as good as or better than before the patch was installed.
To verify these results, we reviewed OS statistics and statspack reports. Here is a typical display from top after the patch was applied:
This clearly shows that the overall load on the machine was much lower and only about 20% of the CPU was utilized. We know from talking to the administrator that users were getting excellent response times even when the system was sending out mass mailings.
We then looked at a statspack report during the same time as the top data above (which was during a mass mailing). We looked at the “Top 5 Wait Events” section first:
We saw that most waits were for I/O and some latching. But, should we have tuned these? As before, these shouldn’t be addressed until we account for the CPU usage. In the excerpt below, we have added the CPU component to the wait components and recomputed the “% Total Time”.
Clearly, sessions are spending most of their time on CPU; if performance were still a problem, we can easily see that I/O events shouldn’t be the main focus of our tuning effort. We know from the top command’s output that even though database sessions are spending almost 89% of their time on CPU, its still a minimal amount of CPU usage overall.
Let’s look now at the SQL statement we tuned (hash value 1656209319):
The statement that we tuned was now responsible for only 2.8% of the total buffer gets and only 2% of the CPU used by the database (21 / 963). We confirmed that the “Gets per Exec” were now only 38.7 compared to approximately 2,600 before the index rebuild. That’s a 99% reduction in logical reads!
Finally, it would be good to see some throughput metrics that convince us that we’re doing at least a comparable amount of work (no tricks up our sleeves). Ideally, the throughput metrics would come from some part of the application itself, but in their absence, we can compare some database metrics:
We focus on metrics such as user calls, transactions, and executions because they imply throughput from the point of view of requests (or actions) per second. bytes sent via SQL*Net to client / sec provides a good metric that accounts for the amount of data returned to clients independent of execution plan efficiency.
Using logical reads / executions is a bad way to compare throughput because it is dependent on how efficiently data is retrieved (i.e., we’re not trying to improve the throughput of our execution plans; we want better plans). On the other hand, bytes sent via SQL*Net to client / logical read will show us how much more efficient the database is after the tuning because (hopefully) fewer logical reads are required for each byte returned back to the client.
From the above metrics we see that the database performed approximately twice as many user calls and transactions after the index rebuild. The logical reads / execution were lower; naturally because now it took fewer logical reads per execution. If we had measured throughput using just logical reads we would have been sorely disappointed (and badly mistaken)!
Our next two statistics clearly show the throughput improvement: approximately 30% more bytes per second sent to the client and approximately six times more data sent for each logical read required to fetch the data (2.58 bytes / read before, 17.87 after).
Keep in mind that all of the throughput improvements came while lowering the overall load on the machine by 95% (load average of 20 down to 1).
In conclusion, we’ve followed the tuning process from the top level symptoms of high CPU consumption at the OS level, down through the symptoms in a statspack report, through a trace of one of the CPU-consuming sessions, finally reaching the point where we were able to tune the query by rebuilding an index. In the end, we were able to confirm the benefit of the query tuning effort by observing significantly reduced CPU consumption, significantly better throughput in the database, and most importantly, happy users!
No comments:
Post a Comment