Diagnosis and Performance Monitoring - Deploying and Managing a Cloud Infrastructure - page 145

5. Consider spreading objects over multiple disks through striping .
6. Determine through calculation the level of performance you can realistically expect
from the system.
The first two steps are just to allow you to understand the extent of the tuning required.
They will also help you identify hot disks in the system. Those are disks that are experi-
encing greater load compared to others. For example, in a 10-disk system, a single disk is
experiencing 25 percent load instead of just 10 percent, which is a good indicator that the
disk is hot.
Analyzing I/O Requirements
If you have a database application running on your cloud system, you should identify key
points in its performance in order to understand which tuning characteristics to implement.
1. Calculate the required and expected throughput of your application so you can estab-
lish a baseline.
You can do this by observing the number of reads and writes involved in each transac-
tion or activity of your application and then identify the objects for which those opera-
tions are being performed. For example, in a single transaction your application might
read from object A, then write to object B, and then do another write to object C. So a
single transaction would yield three I/O operations: one read and two write operations.
2. Define the I/O performance (throughput) of your application or system by specifying
the number of transactions per second (tps) required of the system.
If the system is required to handle 100 tps to perform as expected, then based on the
operations in step 1, the system has to support a total of 300 I/Os per second: 100 reads
to object A, 100 writes to object B, and another 100 writes to object C.
3. Determine the number of disks required to achieve this level of performance.
To do this, ascertain the performance of each disk in terms of the number of I/O opera-
tions it can perform per second. This is usually advertised by the manufacturer, but you
can get a more accurate and actual reading by benchmarking it yourself. This will depend
on three factors:
■ Disk speed.
■ Whether the operation is read or write. Performance varies depending on the
model you have.
■ Whether the file system is being used or just the raw device.
4. Take note in a spread sheet how the disk performs in comparison to the reads and writes
and use that to determine the number of disks required to achieve the performance.
You can create a table similar to the one shown next to tabulate the disk performance
and figure out how many disks are needed for the target performance.