
138 Chapter 7 Optimizing ClusterCATS
Configuring load-balancing metrics
You have the option to customize the load-balancing metrics of web servers clustered
with ClusterCATS software. This section describes how to customize metrics to your
specific website implementation.
Overview of metrics
The JRun and ColdFusion servers record how long it takes to process each request, and
can return metrics derived from this data, upon request. These metrics include the
following:
• Average Request Time (default) the average processing time of all requests within
a one-minute moving window. The use of an average smooths the effects of brief
spikes in request volume, and in a mixture of short- and long-running requests.
• Round Trip Time the round-trip time it takes to process a request.
• Last Request Time (ColdFusion only) the time it takes to process the last request
to the server. Because it is a single, undiluted snapshot of the request time, it
immediately reflects peaks and troughs in request processing time.
To translate these metrics into a single load value for the web server, they must be
weighed against a subjective measure of server performance — a maximum acceptable
response time. The maximum indicates the upper threshold of performance at which a
server is declared “busy” for load-balancing purposes. When a server reaches this critical
threshold, ClusterCATS redirects further service requests away from the server until it
becomes more responsive to its clients.
ClusterCATS software provides a further enhancement in load-balancing options. A
ClusterCATS agent process probes a special page,
getsimpleload.jsp for JRun and
getsimpleload.cfm
for ColdFusion, every five seconds, and records the round-trip time
(RTT) for each request. From this data, it computes its own average RTT over a
one-minute moving window. You can edit this file to customize load-balancing options
for your application.
This external view of request time accounts for the processing time of the request itself,
but, more importantly, for other system overhead included in reaching the web server
and receiving an acceptable response. By factoring in external influences on web server
responsiveness, such as network load, scheduling load, and disk I/O load, the
ClusterCATS probe agent can adjust the load reported by JRun or ColdFusion, to
generate a more realistic picture of the web server's performance for its clients.
For example, if the JRun server reports a light load of requests, but the probe agent finds
significant round-trip times to and from the web server, then it reports a proportionally
higher load for server than JRun reports.
Commentaires sur ces manuels