On 08/05/06, Bennett McElwee <[EMAIL PROTECTED]> wrote:
My results file seems to show a single thread making two simultaneous
requests. This seems to indicate that JMeter is generating bad results.
I wonder if anybody else has seen this?
The following lines appeared in my CSV file for a test. (There were 50
threads total; I have omitted intervening lines for other threads.)
timeStamp,time,label,responseCode,responseMessage,threadName,dataType,su
ccess,failureMessage
1147052818441,360118,Page One,Non HTTP response code,Non HTTP response
message,Thread Group 1-1,text,false,Test failed, text expected to
contain /Testing/
...
1147052854550,72124,Page Two,Non HTTP response code,Non HTTP response
message,Thread Group 1-1,text,false,Test failed, text expected to
contain /Testing/
The second request was apparently initiated at time 1147052854550 -
72124 = 1147052782426. But this is before the first request finished.
This is impossible for a single thread, isn't it? (Note that the "Non
HTTP response" is (I think) some sort of server timeout.)
Yes, that should be impossible, unless the PC clock was changed during
the run - but then other threads would show the same behavior.
Each thread runs one sampler at a time; you can only get overlapping
times for a thread if JMeter is set to download embedded resources,
but in that case the parent sampler would have an elapsed time larger
than any of its children.
The times are generated by calling System.currentTimeMillis() before
and after the sample, and then differencing them. Should be fool-proof
... but I'll try and take a look at the code later.
Are you running on a multi-CPU system?
Might be interesting to see the full set of times - perhaps you can
send the CSV to me privately?
In general I have had anomalous results when testing against a very slow
HTTP server (30-60 second response time) using 50 threads and a constant
throughput of 120 requests/minute overall. I suspect that the ending
timestamps or possibly the times written to the CSV file are wrong, and
the results above seems to prove this. Has anybody else had similar
results?
(JMeter 2.1.1 on WinXP Pro. Nothing unusual in jmeter.log.)
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
