jmeter-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Felix Schumacher <felix.schumac...@internetallee.de>
Subject Re: Important Regression in nightly build compared to 2.13 or r1715087
Date Sun, 31 Jan 2016 17:20:05 GMT
Am 31.01.2016 um 18:07 schrieb Felix Schumacher:
> Am 31.01.2016 um 17:52 schrieb Philippe Mouawad:
>> On Sun, Jan 31, 2016 at 5:45 PM, Felix Schumacher <
>> felix.schumacher@internetallee.de> wrote:
>>
>>> Am 31.01.2016 um 17:41 schrieb Philippe Mouawad:
>>>
>>>> On Sun, Jan 31, 2016 at 5:30 PM, Felix Schumacher <
>>>> felix.schumacher@internetallee.de> wrote:
>>>>
>>>> Am 30.01.2016 um 00:31 schrieb Philippe Mouawad:
>>>>> Hello,
>>>>>> I made a real load test today using nightly build and faced an important
>>>>>> issue.
>>>>>>
>>>>>> Here are the details:
>>>>>> - Test uses 1000 Threads on 1 Instance
>>>>>> - It uses "Download Embedded Resources"
>>>>>> - Socket Timeout is set to 10s
>>>>>>
>>>>>> No time between the requests?
>>>> Variable Time between requests. It is supposed to reproduce user website
>>>> activity.
>>>> The timer is not fixed which can harden comparison but I always reproduce
>>>> the difference of behaviour between version for all runs.
>>>>
>>> So how many requests per second do you simulate (roughly)?
>>>
>> 20 to 30 per second
>>
>>> I tried to load test the ROOT webapp of tomcat 8 (without a timed delay)
>>> and did not get any Exceptions.
>>>
>>> Does it mean there is no network ? only locally ?
>> I am in ideal conditions in term of load testing as I am on a dedicated
>> machine.
> OK. I think I see your problem (even if I don't see the exceptions).
>
> I test a locally running tomcat 8 with 1000 threads running for 1000 
> times one http sampler, which has a gaussian timer that was configured 
> with 30000.0 and 300 milliseconds.
>
> Now, if I run it with httpclient4 I get errors after a short period of 
> time (about 5 to 10 seconds). No errors, when run with httpclient3.

If I let the test run with one round, only, there is no error. The same 
is true, when I disable keep alive. So it seems to be a change in keep 
alive handling.

Regards,
  Felix
>
>
> Regards,
>  Felix
> ||
>>> Regards,
>>>   Felix
>>>
>>>
>>>
>>>> Is your server capable of serving thousand requests simultaneously?
>>>>> Yes, no problem on this side.
>>>> Number of requests per second is not high at all (50 samples included
>>>> Transaction Controller which encapsulate Http Requests ).
>>>>
>>>>
>>>> Regards,
>>>>>    Felix
>>>>>
>>>>> There is no overloading of the machine, no impacting GC
>>>>>> Very rapidly, I start getting a lot of errors:
>>>>>>
>>>>>>
>>>>>>       - Non HTTP response code:
>>>>>> org.apache.http.conn.ConnectTimeoutException
>>>>>>       message:Non HTTP response message: Connect towww.foo.com:80
 timed
>>>>>> out
>>>>>>
>>>>>>
>>>>>> Rate of error varies between 15% and 30%.
>>>>>> Note that if I navigate on the application, I don't face the errors.
>>>>>>
>>>>>> I ran the same test using exactly the same configuration:
>>>>>>
>>>>>>       - Same machine
>>>>>>       - Same JVM version and tuning
>>>>>>       - Same user.properties
>>>>>>       - Same hc.parameters
>>>>>>
>>>>>> But  jmeter  r1715087
>>>>>> And error rate is 0.30%.
>>>>>>
>>>>>>
>>>>>> Note the target server has a load balancer that returns a keep-alive
>>>>>> duration set to 2 (2 seconds).
>>>>>>
>>>>>> This issue is a blocker one for the release of next version.
>>>>>> I compared code with revision 1715087 and I don't see many changes
in
>>>>>> HTTPHC4Impl that would explain this regression.
>>>>>> I commented out some suspects , retried but I get same results.
>>>>>>
>>>>>>
>>>>>> I also upgraded to HttpClient 4.5.2 and uncommented the code expected
to
>>>>>> be
>>>>>> added, same results.
>>>>>>
>>>>>> So for now I tend to suspect an issue in HttpClient/Core.
>>>>>>
>>>>>>
>


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message