jmeter-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From sebb <seb...@gmail.com>
Subject Re: Design of Timeout test element / sampler interrupter
Date Sun, 30 Aug 2015 15:36:23 GMT
On 30 August 2015 at 11:54, sebb <sebbaz@gmail.com> wrote:
> On 30 August 2015 at 07:12, Philippe Mouawad <philippe.mouawad@gmail.com> wrote:
>> On Sunday, August 30, 2015, sebb <sebbaz@gmail.com> wrote:
>>
>>> I've had a look at the classes that implement SampleListener, and
>>> apart from ResultAction and TransactionSampler, only the Listeners use
>>> it. Since usage of these should be minimised in a production test,
>>> it's likely that there won't be as many implementations as I had
>>> feard.
>>
>> Implementation of SampleListener ?
>> Usage of which should be minimised ?  SampleListener ?
>
> Usage of additional  Listeners should be minimised in a production test.
>
>>>
>>> Also if the implementation is empty, the overhead will be quite small.
>>>
>>> [There is a work-round if it does prove expensive: the SampleListener
>>> interface could be split into two parent interfaces.]
>>>
>>> So assuming that JMeterThread implements sampleStarted/sampleStopped,
>>> the Timeout element can use the Start to set up the timer and the Stop
>>> to cancel it. This will reduce the number outstanding as much as
>>> possible.
>>>
>>> The timeouts have to be implemented using separate threads for two reasons:
>>> - it's obviously not possible to interrupt a sampler from the same
>>> thread as the sampler
>>> - depending on the sampler, and its state, the interrupt may take a
>>> while to complete, so each interrupt must be done in its own thread
>>
>> Are you sure, calling interrupt is usually just about setting a flag no?
>
> Ultimately yes, I guess a flag will be set.
> However I'm not sure that it is always instantaneous, as there may be
> locks to aquire.
>
>> Having 1 thread for each interruption, could lead to hundred of
>> threads running for high throughput threads (500 res/s for example), it
>> won't scale.
>
> That assumes that all the samples in all the threads have timeouts enabled.

It also assumes that the threads are running for the life of the
timeout, which I've just realised is not the case.
>> Why can't we have 1 Thread (TimeoutChecker) called every N milliseconds
>> that checks all registered JMeterThreads to check and call interrupt if
>> necessary ?
>>
>
> That's quite a bit of work to code, but if there is already such a
> queing mechanism it would be worth trying.

Actually the ScheduledExecutorService does just that.
It uses one thread to handle the timeouts in sequence, and then starts
a new Runnable when the timeout expires.

Now that I have created the service as a singleton, there will only be
one extra thread (the executor) most of the time.
Since the timer tasks are cancelled if the sample completes in time,
there will only ever be at most one extra thread for each overdue
sample.

I think it is a reasonable assumption to assume that the number of
such samples will generally be small.
At the very most it could only be one per thread.
And remember that the new thread is only started when the timeout expires.
It is extremely unlikely that they will all expire at once, and anyway
(according to you) they don't take long to run so they won't build up.

I've just noticed that there is a version of the Executor Service that
uses Callable rather than Runnable (not sure why I missed it before).
I asume this means the tasks are run in the same thread (I'll check this).
If the interrupts really do happen quickly, this might be a better choice.
Even if not, then it may not matter so long as any delays don't
continue to build up.

>>
>>> It should be possible to use a single shared instance of the
>>> ScheduledExecutorService; that could be lazily created using IODH. [I
>>> can try that with the current implementation]
>
> I've already tried it.
>
>> See my note above
>
>>
>>>
>>> As to whether the Timeout class should be a Timer or some other type
>>> of test element - that does not matter so long as it can be applied to
>>> the samplers individually or when in scope.
>>>
>>> I chose Timer because it was already called in the right place, but I
>>> assume JMeterThread can call any Test class provided that it
>>> implemented the SampleInterface.
>>>
>>> It must be one of the existing Test Element classes that are handled
>>> by the Menu system otherwise it will need special handling.
>>>
>>> The scope requirement rules out Config elements and Logic Controllers.
>>
>> It does not seem like a pre-processor to me, nor a post-procesor, nor a
>>> Listener
>>>
>>> So AFAICT the only remaining options are the Timer and Assertions.
>>>
>>> I think both are justifiable.
>>
>> Why isn't it part of Sampler abstract class and as such a field in Sampler?
>
> How does the user indicate that a Timeout should be applied to a
> particular Sampler?
> It's a lot of work to add Timeout fields to every GUI.
> Whereas being able to add a child test element to each applicable
> sampler is already supported.
> Further, such a test element can be applied to multiple samplers in scope.
> Much easier to enable and disable a single element that having to
> update each Sampler.
>
> That's why it needs to be a separate test element.
>
>>
>>
>> For me none of Timer nor Assertions are conceptually valid.
>> The behaviour is not a pause(so not Timer for me), it's not an Assertion
>> neither as for me an Assertion only checks something.
>> Although not fully satisfying it look to me more of a PreProcessor as it
>> sets a timeout on the Sampler , it can also be considered as a Post
>> Processor.
>
> OK, I could live with it being a Pre-Procesor.
> That has the correct scoping rules.
>
>>
>>
>>>
>>> The name of the class can of course be changed from InterruptTimer - I
>>> think that is probably not the best choice. Maybe something like
>>> SamplerTimeout?
>>>
>> Yes or SamplerTimeouter or SamplerInterrupter
>>
>
> Or even SampleTimeout - that's what it does, it applies a timeout to a sample.
>
>> --
>> Cordialement.
>> Philippe Mouawad.

Mime
View raw message