Demo of Percentile Metrics
Percentile metrics are available for business transaction response time. Percentiles are generally a better indicator than averages because they are not sensitive to outliers while averages can get distorted. Percentile metrics can provide a sense of how the response times are distributed. From this information you can extrapolate the percentage of responses that are within and outside of acceptable ranges.
Percentiles describe distributions of real world data sets in ways that are less sensitive to the effect of outliers in the data set than simpler calculations such as the mean. Percentile metrics can provide you with answers to questions such as the following:
- What is the 95th percentile latency of transactions for a single web server?
- What is the 95th percentile latency of transactions for the entire web site (over all the servers)?
- What is the 95th percentile latency of transactions for the website during the last one hour?
Percentile metrics for business transaction response time are displayed in the Metric Browser or Custom Dashboard.
Percentile metrics can be interpreted as follows: the 95th percentile point indicates that 95% of all response times were less than the metric value. It provides a sense of how the response times are distributed. For example if the metric Observed (Average) or percentile value is 300ms, it implies that 95% of the business transaction response times are less than 300ms, and therefore implies that 95% of the business transaction response times are within acceptable ranges. It also implies that 5% of the requests have been taking more time and could be a cause for concern.
Enabling Percentile Metrics
There are three agent node properties that you can use to enable and configure how percentile metrics are collected. You do not need to restart the agent after making changes to these properties.
- disable-percentile-metrics: Set this property to 'false' to view the percentile metrics.
- percentile-method-option You can choose one of two different algorithms to calculate percentiles in AppDynamics:
- P Square algorithm (default): This option consumes the least amount of storage and incurs the least amount of CPU overhead. The accuracy of the percentile calculated varies depending on the nature of the distribution of the response times. You should use this option unless you doubt the accuracy of the percentiles presented.
- Quantile Digest algorithm: This option consumes slightly more storage and CPU overhead for the machine where the agent is running, but may offer better percentiles depending on how the response times are distributed.
- percentiles-to-report: By default, the system will capture the 95th percentile metrics. You can change the percentile captured here.