We must remember that Oracle does not run in a vacuum, and that there are important "external" events that take place outside the scope of the Oracle instance.  We see this frequently with the Oracle disk I/O metrics.

I/O remains a critical part of Oracle performance tuning.  While everything else in Oracle runs in microseconds and RAM moves in nanoseconds, the mechanical disk devices on some Oracle servers become a major bottleneck.

When tuning Oracle, we often forget to look outside the instance, examining external influences on disk I/O speed.  These may include:

• Stripe size - The stripe size for raw disk partitions will influence I/O behavior, especially for multi-block read operations (full scans).
   

• TCP/IP - TCP IP settings and Oracle TNS settings (e.g. tcp.nodelay) effect I/O timings.
   

• RAID level - The RAID used on the disk array can have a profound impact on end-to-end I/O timings.  For example, up until 2015, RAID 5 with Oracle was sometimes inappropriate for high-update databases. 
   

• Disk Controllers - A bottleneck in the disk farm such as a lack of controller resources can precipitate high I/O latency.
   

• Storage Array Internals - Many of today's storage arrays have specialized optimization software and on-board RAM caching to improve throughput.
   

• SAN/NAS - Dynamically attached storage has special issues.

• Server caching - Servers have both an internal file cache vs. JFS cache

When Oracle makes a physical I/O request, it's handed-off to the operating system as a native I/O operation.  At this layer, the device-media interface takes over, as Oracle patiently waits for the block to be returned to the calling routine.  During this time, many external events can influence the observed I/O timings.

Disks reads and duplicate RAM layers

As hardware evolved though the 1990's, independent components of database systems started to employ their own RAM caching tools as shown below:

In this figure, the Oracle database is not the only component to utilize RAM caching.  The disk array employs a RAM cache, the servers have a Journal File System (JFS) RAM cache, and the front-end web server also serves to cache Oracle data.
 
This concept is important because many enterprises may inadvertently double cache Oracle data.  Even more problematic are the fake statistics reported by Oracle when multiple level caches are employed:

• Fake Physical I/O times: If a disk array with a built-in RAM cache is in use, the disk I/O subsystem may acknowledge (?ack?) a physical write to Oracle, when in reality the data has not yet been written to the physical disk spindle.  This ?false ack? can skew timing of disk read/write speeds.
   

• Wasted Oracle Data Buffer RAM: In system that employs web servers, the Apache front end may cache frequently used data.  Thus, significant Oracle resources may be wasted by caching data blocks that are already cached on the web server tier.
   

Disk writes and false "acks"

Many of today's disk arrays employ a delayed write mechanism, improving performance by caching the data block in RAM, and writing it asynchronously.  At write time, these devices lie to Oracle (by "acknowledging" the disk write with an "ack"), when in-reality the data block is still in RAM on the storage array.  This will skew I/O timing, reporting them as lower values than a completed disk write.

Let's take a closer look and see why external influences can bias performance tuning studies.

Oracle metrics and disk performance

As we have noted, the reported times that Oracle gets back from the OS can be exaggerated (the false "ack"), and the root cause of slow I/O is frequently hidden from Oracle by the machinations of the operating system.

This issue is also aggravated by RAM caches (Disks cache, JFS cache) which hide the "real I/O timings.  See my notes on multiple RAM levels for complete details.

Anytime an Oracle data block is accessed from disk, we commonly see three sources of delay. The first and most important source of delay is the read-write head movement time. This is the time required for the read-write head to position itself under the appropriate cylinder.

We also see rotational delay as the read-write head waits for the desired block the past beneath it, and the third source of delay is the data transmission time from the disk back to the Oracle SGA.

Low seek time improves throughput
(place hot files near middle absolute track)

If we accept the premise that 99 percent of the latency is incurred prior to actually accessing the desired data block, then it makes sense that the marginal cost for reading a 32K block is not significantly greater than the cost of reading a 2K block. In other words, the amount of disk delay is approximately the same regardless of the size of the block. Therefore it should follow that the larger the block that you can read in on a single I/O, the less overall I/O will be performed on the Oracle database.

The principle behind caching is not unique to Oracle databases. Access for RAM is measured in nanoseconds, while access from disk is generally measured in milliseconds. This amounts to a to an order of magnitude improvement in performance if we can get the Oracle data block into a RAM buffer.

As Oracle grows more sophisticated and RAM becomes cheaper, we tend to see Oracle databases with system global areas (SGA) that commonly exceed 10 GB. This has important ramifications for the performance of the Oracle database because once read, the Oracle data blocks reside in RAM where they can be accessed tens of thousands of times faster than having to go to disk in order to retrieve the data block.  For details, see my notes on disk I/O concepts.

Perception vs. reality in disk I/O

Consider this compelling argument that different block sizes should have no influence on total response time.  They correctly note that, in theory, four 8k block reads should take the same amount of time time as a single 32k read, and ergo, that block size should not matter.

However, this does not explain credible reports of different response times using different block sizes.  Could these "external" influences be to blame?