Diagnosis and Performance Monitoring - Deploying and Managing a Cloud Infrastructure - page 131

Cooperative Caching
Processing performance increases more rapidly than disk performance, and that is one
reason the file system should use caching to avoid frequent disk access. But another
technology trend is also prompting the benefits of caching. Current and emerging high-
speed low-latency switching networks enable the supply of file system blocks across
network spaces much faster than standard Ethernet. This means that fetching data from
remote memory is faster than ever, signaling a good performance boost for distributed
file systems, NFS, and remote virtual machines that are at the core of cloud computing.
Applications and file systems are now able to cache data using not just local memory and
resources but that of networked resources as well. This prevents the issues that plague
write backs if ever there are unexpected crashes.
One example of a cooperative caching algorithm is Direct Client Cooperation. This is
a simple approach that allows any active client to use another idle client's memory as a
backing store. An active client with an overflowing cache can forward new cache entries
to an idle machine. The active client can use this remote private cache for read requests
and other caching applications until that idle client becomes active and needs its cache
back, at which point it evicts the cooperative cache. This is a good method to use in a
multitenant environment because not all virtual machines will be available all the time.
This helps lighten the load for everyone and does not require the provisioning of more
resources for a single instance.
Bandwidth
Aside from the performance of core low-level cloud technologies discussed previously,
the most important aspect for a cloud system is the actual bandwidth available. It is a well-
known fact that unpredictable network performance is a major factor that affects user
satisfaction of cloud services. No matter the performance of the cloud system, if the receiv-
ing end is a bottleneck, the perception of overall performance will be low. If there is no net-
work performance guarantee, there is also no application guarantee. But since the Internet
is a collection of globally shared systems from multiple vendors and providers, network
quality and hence bandwidth cannot be guaranteed.
Bandwidth and overall network performance are often the cause of major user com-
plaints. It is not unusual for administrators and data center managers to be brooding
over a collection of complaint tickets expressing user frustration because of application