In LoadRunner, the percentile graph is a great graph for understanding how well the system is performing in terms of transaction response time from an overall perspective in a load test.  How is the percentile graph derived?   The mathematical explanation/formula can be found in Wiki, Percentile rank. 

90% percentile allows us to know that for a certain amount of transactions are performing at a certain amount of duration.  Of course it's a really abstract description but anyway how is it useful?  Example, if the 90th Percentile is 10secs, what it literally means is that 90% of all transactions are performing at or under 10secs.  While the remaining 10% are performing more than 10secs.  If you have a SLA or performance requirements that is based on 90% of the transactions to perform under 8secs, you have pass this criteria. 

The Percentile can also tell you how the system is performing, such that if there is a higher percentage of slow response time transaction and what is the percentage that is performing higher than the expected transaction response time.  In a typical average transaction response time graph, you can see how the (passed) transactions are performing during the entire load test period over time.  However, the Percentile graph illustrates the (passed) transactions and their percentage with respect to the overall transactions. 

In the below example taken from the Analysis User Guide, at the 60th Percentile, tr_amazon_help is at about 4secs, which means 60% and less of tr_amazon_help is responding at 4secs less.  While at the 90th Percentile, 90% of all tr_cnn_weather transactions are responding at 20secs or less.