Capacity Planning - The Practice of Cloud System Administration - page 381

haveavailable.Tryseveraldifferentcombinationsofshortandlongperiods,andseewhich
combination best predicts trends observed in your historical datasets.
18.2.6 Monitoring the Key Indicators
The correlation coefficients between core drivers and resources should be graphed and
monitored. If there is a significant change in correlation, the relationship between the core
driver and resource consumption should be reassessed, and any changes fed back into the
capacity planning process.
Similarly, the MACD and MACD signal for each core driver and resource should be
monitored. Since these metrics can be used for early detection in changes in trends, they
arekeyelements incapacity management. AlertsgeneratedbytheMACDlineandMACD
signal line crossing should go straight to the team responsible for capacity planning. If
the core drivers and relationships to the primary resources are well defined, theoretically
it should be necessary to monitor only the core drivers and the correlation coefficients. In
reality, it is prudent to monitor everything to minimize the chance of surprises.
18.2.7 Delegating Capacity Planning
Capacity planning is often done by the technical staff. However, with good metrics and
a clear understanding of how core drivers affect your resource requirements, you can de-
couplethecapacityplanningfromthedeployment.Aprogrammanagercandothecapacity
planning and ordering, while technical staff take care of deployment.
You can enable non-technical staff to do the capacity planning by building a capacity
planning dashboard as part of your monitoring system. Create one or more web pages with
capacity data in a specialized view, ideally with the ability to create graphs automatically.
Make this dashboard accessible within the organization separately from the main monit-
oring dashboard. This way anyone in the organization can access the data in a reasonable
form to justify capital expenditure on additional capacity.
Sometimes the decision is not to buy more resources but rather to make more efficient
use of existing resources. Examples might include compressing data rather than buying
more storage or bandwidth, or using better algorithms rather than increasing memory or
CPU. Testing such scenarios relies on resource regression, which we will examine next.