Can't make Avg90 work out

[Chrondeath 0]

I'm using WAPT 2.5.

 

I thought I understood how the Avg90 number was supposed to work: 90% of the response times were under this value. I was thinking that means "sort all the response times, go 90% of the way through that list, and the avg90 number should be slightly higher than that entry."

 

On our latest run, while looking at the report, one of the blocks of data for one type of request read like this:

 

Total requests: 216

Min: 0.12(0.12)

Max: 10.1(10.1)

Avg: 0.19(0.19)

Avg90: 1.06(1.06)

 

I had a co-worker express skepticism at the Avg90 number there. Our current main performance problem suspect is the Java garbage collector, so we weren't surprised to see a few instances where the garbage collector kicked in during the request and it took ten seconds, but the rest of the time things seemed to be consistently under a second. If our times were all either sub-second or 10+ seconds, how could Avg90 be 1.06 seconds?

 

So I took a look at the CSV file of timings that WAPT generated, and I can't come up with that Avg90 number. Subtracting "Request Begin" from "Response End" for every line of that profile entry (not just the 216 for that chunk), the six highest overall were:

 

0:00:00.328000
0:00:00.359000
0:00:00.372000
0:00:00.404000
0:00:00.448000
0:00:10.111000

 

How can there be an avg90 of 1.06 seconds in any chunk if only one request ever took more than one second?

 

I was able to compute numbers that matched the min, max, and average, so I don't think I've screwed up parsing the file.

[sergei 0]

We use standard deviation to calculate Avg90. But yes, this method is not always correct.