In order to respond to a request, servers utilize a variety of resources such as databases, network bandwidth, system memory and CPU. The server must hold many of these resources until the required response is prepared and then transmitted back to the client. The fulfillment of a request therefore involves network delays and some processing time on the client side as it interprets the response it receives.
This metric measures the total time needed to build a response for the request and send the results back to the client. During that time, many of the resources required to formulate the response may be partially or fully dedicated. The following delays are included in this metric:
Some commands include a request but not a response because the user halts the command (with a TCP Reset), or because a server is overloaded. For these commands, the command completion time is undefined and is not included in the calculation for this metric.
This metric is incremented with each new request the user sends to the server, even if the response is never received. However, the metric is not updated until the response is completed or the system determines the response is never sent (connection reset or connection time out).
This metric reflects the time required to process the request, build a response and begin the initial transmission of the response. This metric does not include any network delays or client-side processing delays.
Some commands include a request, but not a response, due to the client halting the command (with a TCP Reset) or due to the server being overloaded. For these commands, the command response time is undefined and is not included in the mean, minimum, maximum, or standard deviation statistics.
Percentile metrics provide a means to quantify the quality of service customers are receiving based on the response time that a fixed percentage (usually 95 percent) of clients receive. These metrics are commonly referred to as 95th Percentile metrics. The exact percentage the system uses is configurable by the user. This method has the advantage over relying on the arithmetic mean in that it excludes outliers which can distort the mean.
To calculate this metric, the system uses a statistical method involving standard deviation rather than tracking each individual event in memory (that approach would be prohibitive in terms of processing requirements and memory consumption). The percentile metric is calculated using this formula: