Oracle & read ahead cache Oracle Database Tips by Donald Burleson

Question:  I just wanted to check with the experts on "Read ahead cache" on Windows 2003 Sp2 SAN configuration.  Currently my SAN doesn't employ "Read ahead cache" turned on because the nature of the application is high volume very random selects and huge amount of inserts.    Here are my top-5 events from an AWR report:

Top 5 Timed Events
                                                     Pct
Event                          Waits   Time (s) (ms) Time
db file sequential read        445,254 2,735     6 70.0
CPU time                                       650 16.6
enq: TX - row lock contention       95 279    2933  7.1
db file parallel write         421,399 125       0  3.2
db file scattered read         133,905  90       1  2.3

Under what conditions will a enabling the read ahead cache help me?
 

Answer: The read ahead cache is only useful in cases of "db file scattered reads" (i.e. full table scans).  In your example, you only have 2.3% of I/O as scattered reads, and a read-ahead may not help much.

The read ahead caching has many names on different operating systems, and it started on IBM mainframes as "sequential prefetch".

The concept of readahead caching is simple.  Over 90% of I/O latency is consumed in the read-write head movement, as the heads are placed under the target cylinder.  Once in-place, the disk continues to spin and the read-write head can simultaneously transmit back the original block request at the same time as the next sequential block passes below the read-write heads.

For complete details on device-level caching, see my book "Oracle Tuning: The Definitive Reference".

For scan operations (index range scans, index fast full scans, and full-table scans), a read-ahead cache can be very useful for speeding up these "scattered read" operations.

In sum, if your database is requesting adjacent data blocks, the read-ahead cache may improve I/O throughput.  However, using RAID10 (SAME, stripe and mirror everywhere), like with ASM, the blocks are not adjacent, and a read-ahead cache may not be as useful as a database where the data blocks are laid-out sequentially.  Of course, the stripe width influences this decision.

For complete details, see the book "Oracle disk I/O tuning" where we see specific examples.

Placing adjacent blocks together on disk (when you know that they will be accessed sequentially) is analogous to the process of placing high-impact files near the middle absolute track of huge Oracle disks, to minimize read-write head latency:

Placing hit impact files in the middle cylinder reduces read-write head movement.

How to tell of read-ahead caching is right for your Oracle database

As we noted, enabling your read-ahead cache will greatly improve the I/O throughput for scan operations.  In a normal database you should look for "db file scattered reads", and see if the latency is less than "db file sequential reads", which happen on single block fetch operations:

Using Ion to find disk I/O latency

The Ion tool is the easiest way to analyze STATSPACK and AWR disk I/O data in Oracle.  Ion allows you to spot hidden I/O trends and times when full-scan operations are important.

In sum, a transaction that does not require adjacent data block will not benefit from a read-ahead cache, while transactions that perform index range scans or full-table scans will benefit the most from a read-ahead cache.

Also, remember that most Oracle database process a wide range of transactions and an OLTP database which does few "db file scattered reads" may perform lots of full-scans during batch reporting windows.

Hence, you need to compare the costs of enabling your read-ahead cache with the benefits in speeding-up "db file scattered reads".  This can easily be confirmed in a STATSPACK or AWR report.

 

If you like Oracle tuning, see the book "Oracle Tuning: The Definitive Reference", with 950 pages of tuning tips and scripts.  You can buy it direct from the publisher for 30%-off and get instant access to the code depot of Oracle tuning scripts.