Diagnosis and Performance Monitoring - Deploying and Managing a Cloud Infrastructure - page 143

Amazon categorizes its EC2 instance types ( http://aws.amazon.com/ec2/instance-
types/ ) with various storage offerings into general purpose, compute optimized, memory
optimized, and storage optimized. Amazon basically makes available two types of main
storage devices: hard disk drives (HDDs) and solid-state drives (SSDs). These come with
different performance levels and costs.
An HDD will be generally suitable for batch processing systems or systems that do not
need to be used in real time. These include large databases (SQL or NoSQL), log storage,
recorded audio, and video streaming. SSDs will be highly relevant for web servers serving
millions of page hits per day. Moreover, data crunching systems, especially those involving
GPUs, will definitely benefit from SSDs as compared to HDDs.
Disk Tuning
Disk tuning is the optimization of a disk's performance in relation to the system or platform
that it has been installed on, taking into consideration the operating system and the file sys-
tem being used.
The traditional approach to disk tuning is to use the advertised performance characteristics
of the disk such as those discussed previously and then to create performance expectations.
You can then benchmark the device and compare the actual performance to the modeled set of
expectations.
Disk performance characteristics are largely consistent when you consider only the disk
itself. What performance differences you do get from them is largely due to the kind of file
system being used and how it manages the disks as well as I/O performance. Aside from the
file system performance affecting disk performance, one very good way of maintaining disk
performance is regular defragmentation. Fragmentation occurs simply because the processor
and other electronic components of a computer system are leagues faster than the mechanical
disk. To compensate, the disk will write to the free sector that is closest to its actuator and
head; it will not seek out large empty portions to write large contiguous files into. Because of
this, the disk eventually becomes fragmented, and parts of a single file are all over the place.
Performance slows down because when a file is needed, the disk has to seek out those frag-
ments and then piece them together, accumulating overhead because of all the disk seek and
access being done to collect each fragment. The problem eventually exacerbates, and you are
left with a system that crawls whenever an I/O operation is in progress.
Fragmentation typically occurs on logical drives, which are translated by the device
drivers into the physical disk. In the case of virtual systems, as is very common in cloud
computing, the logical volume, which is already a masking of the physical medium, is fur-
ther masked by another application layer, the virtual disk. Virtual disks reside on logical
disks as container files and as such are still prone to fragmentation in the logical drive level.
Further, the virtual disk itself, perfectly emulating a physical disk, also gets fragmented.
Now you have a double fragmentation scenario. The virtual disk is fragmented, the frag-
ments are mapped onto the logical drives, which in turn are mapped on the physical drive.
This would explain various performance issues in virtual machines.
Combating this is a simple process using a tool that comes mostly standard to any
operating system, a disk defragmenter. Logical drives and especially virtual drives that