There are many Oracle experts who have very different approaches to Oracle tuning and database optimization. There are Oracle scientists, pragmatists and statisticians, all of whom are successful at Oracle tuning, but with very different approaches
The problem occurs when some Oracle experts become zealous about their method and a "dogma" emerges. Some Oracle guru's even develop a cult-like following, each chanting the mantras of their opinion leader and fostering an almost religious fervency in their belief system about Oracle tuning. Cults are fine, but in the real-world, we use every approach that has validity!
Many Oracle gurus' publish "rules" for the best way to tune an Oracle database, and Oracle professionals accept these rules as Gospel. This acceptance of rules is because of their high respect for the Oracle guru, but it is also a function of the needs of a tuning professional to have "firm" guidelines and checklists to follow. If only it were that simple! Of course, these rules are often little more than tautologies and they only serve to confuse the Oracle community.
In reality, Oracle tuning success is a semi-structured task, and the successful tuner must have an excellent understanding of all of the interacting components to be successful. There are no hard-and-fast rules; everything is in flux.
Let's take a close look at some of these Oracle tuning fallacies and understand their weaknesses.
Fallacy: The sole purpose of Oracle tuning should be to reduce logical I/O
This is a dangerous fallacy because it ignores important external factors such as disk, network and CPU bottlenecks and internal factors such as the library cache. While reducing logical I/O will minimize impact on the data buffer cache, there is no guarantee that physical disk I/O will be reduced.
Accessing Oracle data via the data buffers is at least 10x faster (sometimes hundreds of times faster) than disk access, so one should not ignore disk enqueues and assume that only tuning-down logical I/O will remove the problem. Oracle created the KEEP pool so that the DBA can identify and cache high-use tables and indexes.
Fallacy: Hourly tuning averages are meaningless
Some Oracle scientists advocate taking point-in-time samples of performance and argue that any elapsed-time average of Oracle performance longer than 15 minutes is too long to be valid. However, many shops recognize that hourly STATSPACK and AWR metrics (average buffer-hit-ratio, library cache miss ratio) are often critical for long-term proactive tuning.
For example, collecting hourly STATSPACK or AWR averages over many months allows the DBA to develop "signatures" of Oracle performance. These signatures will clue the Oracle DBA to times when their system can be proactively reconfigured to anticipate changes in processing modes. I know of one shop that detected a hidden batch job every Tuesday evening, and they were able to cut the run duration in half by reconfiguring the SGA.
Fallacy: Only transaction-level tuning will be successful
Some Oracle tuning professionals collect Oracle 10046 trace dumps and use these cryptic hex dumps top locate poorly-performing SQL statements. They shun the system-level tuning approach as treating the "symptom" and not the cause of the performance problem.
They argue that if they tune at the application and SQL level, then the system-wide issues will self-optimize. In reality, both transaction-level and system-level tuning are critical activities. Of course you must optimize your SQL to ensure that it is gathering the result set with the minimum amount of logical I/O, but the overall configuration is also critical. I have made a single change to an Oracle init.ora parameters and seen the performance of hundreds of SQL statements improve.
System-level tuning is a reality. Some tuning professionals treat the system-wide issues first and then drill-in to the sub-optimal PL/SQL and SQL. Other shops that do not have the large amount of funds required to pay for detailed application-level changes (tweaking application code & SQL). Here are some examples of real-world system-level Oracle tuning activities:
- Replace sub-optimal RAID - The wrong RAID level for the Oracle application can be disastrous. For example, write-intensive Oracle databases (high DML) will often perform faster with RAID 0+1 than RAID5. For details, see the Battle Against Any RAID Five at www.baarf.com.
- Fixing disk bottlenecks - High waits on background events like "db file parallel write" and "log file parallel write" are great indicators of disk enqueues. If your top wait event is relates to files reads ("db file sequential reads" and "db file scattered reads") then you may want to check for disk enqueues and see if striping (RAID 0+1, RAID10, Stripe and mirror Everywhere SAME) might relieve the bottleneck. Also some shops move high impact data files onto solid-state disks.
- Change sub-optimal optimizer parameters - Resetting a sub-optimal setting for several optimizer parameters (optimizer_mode, optimizer_index_cost_adj) can double the performance of some systems. Changing optimizer_mode=rule to optimizer_mode=first rows (or vice-versa sometimes!) can influence the performance of every SQL statement in the application.
- Fix poor-quality optimizer statistics - The Oracle cost-based SQL optimizer (CBO) will only make good decisions if it has good-quality samples from the schema objects.
- Reconfigure the SGA - The setting of the SGA region sizes and sub-area will have a major impact of overall performance.
- Use a faster network - System where the top Wait Event is consistently related to sqlnet waits then you can add faster network resources. However, remember that many network bottlenecks can be fixed by tuning the application to make less Oracle calls, or by adjusting the TNS parameters (sqlnet.ora, tnsnames.ora), specifically the tcp.nodelay, SDU and TDU parameters.
- Get faster processors - Faster (or more) CPUs may greatly improve performance of Oracle system that are constrained by CPU (as determined by Oracle Wait Event reports). Theses types of systems may perform up to 10x faster depending upon the speed and SMP capabilities of the server. The latest trend is to run Oracle on the new 64-bit Itanium2 servers using Windows or Linux.
Fallacy: If an Oracle tuning metric has an exception, it is not valid
Many scientist Oracle professionals thrill when they find that a general rule of Oracle performance can be challenged.
For example, showing how the data buffer cache ratio can be manipulated at-will does not negate the fact that RAM caching (and the ratio's to monitor caching) are a critical part of Oracle tuning. Data blocks in RAM can be accessed in speeds that are several orders-of-magnitude faster than traditional disk access.
One of the most obtuse arguments I've heard in quite awhile is that a high buffer cache hit ratio (95%) may be symptomatic of a poorly-tuned database I/O subsystem because of sub-optimal SQL that is causing high logical I/O! Of course, a high buffer cache hit ratio may also signify a well-tuned database.
A high buffer cache hit ratio does not really tell you much, but a low buffer cache hit ratio for an OLTP database almost always indicates that the buffer is too-small to cache the working-set of frequently-referenced blocks. In fact, ratio's form the basis for the Oracle10g AMM facility and the Oracle v$db_cache_advice utility.
We must remember that all Oracle tuning activities are measurable, and therefore can be proven successful. However, the approach for finding the problem is very personal, and many tuning professionals keep a vast array of tools and approaches at their disposal.
Beware of any Oracle professional who publishes hard-and-fast rules for Oracle tuning. There is no shortcut to successful Oracle tuning, and the tuner must intimately understand the complex interactions of the Oracle subsystems.
About the Author
Donald K. Burleson has been a DBA for more than 20 years and provides Oracle consulting for
systems that require high performance. The author of more than 30 books, Burleson provides Oracle
consulting at www.dba-oracle.com and remote Oracle support at www.remote-dba.net. If you like
Oracle tuning, you might enjoy my latest book "Creating a Self-tuning Oracle Database" by Rampant
TechPress. It's only $9.95 (I don't think it is right to charge a fortune for books!) and you can
buy it right now here.
This was first published in September 2004