Hey saith,

This is great. So when you removed the duplicate windows, there were no OOM
issues?

Seshi
On 24 Jun 2016 14:28, "Sajith Ravindra" <[email protected]> wrote:

> Hi Malith,
>
> Thanks for the explanation.
>>
>> I would expect some variation in the throughput. The aim should be
>> minimize the variation in the throughput (while maintaining the throughput
>> at its highest level).
>>
>  Agreed. Actually, our expectation is to minimize the fluctuation and to
> increase the throughput.
>
>>
>> It is possible to measure the latency as well?
>>
> In IS-Analytics we don't generate any output events or alerts currently.
> Therefore, we don't calculate the latency. What we do is summarize events
> using siddhi queries and persist them in a DB and further summarize
> persisted data using Spark and then displayed through dashbaords.
>
>
> Thanks
> *,Sajith Ravindra*
> Senior Software Engineer
> WSO2 Inc.; http://wso2.com
> lean.enterprise.middleware
>
> mobile: +94 77 2273550
> blog: http://sajithr.blogspot.com/
> <http://lk.linkedin.com/pub/shani-ranasinghe/34/111/ab>
>
> On Fri, Jun 24, 2016 at 1:57 PM, Malith Jayasinghe <[email protected]>
> wrote:
>
>> Hi Sajith,
>>
>> Thanks for the explanation.
>>
>> I would expect some variation in the throughput. The aim should be
>> minimize the variation in the throughput (while maintaining the throughput
>> at its highest level).
>>
>> It is possible to measure the latency as well?
>>
>> regards
>>
>> Malith
>>
>>
>>
>> On Thu, Jun 23, 2016 at 11:13 PM, Sajith Ravindra <[email protected]>
>> wrote:
>>
>>> Hi Malith,
>>>
>>> The deployment is just a standalone DAS server, and we are planning to
>>> do a test for HA deployment in recent future.
>>>
>>> The workload is generated by a .csv data file which has 100K sample
>>> events, 10M events are generated by iterating through the same data set 100
>>> times. But we keep increasing the timestamp. A simple thrift client is used
>>> to publish data.
>>>
>>>
>>>
>>> Thanks
>>> *,Sajith Ravindra*
>>> Senior Software Engineer
>>> WSO2 Inc.; http://wso2.com
>>> lean.enterprise.middleware
>>>
>>> mobile: +94 77 2273550
>>> blog: http://sajithr.blogspot.com/
>>> <http://lk.linkedin.com/pub/shani-ranasinghe/34/111/ab>
>>>
>>> On Fri, Jun 24, 2016 at 10:34 AM, Malith Jayasinghe <[email protected]>
>>> wrote:
>>>
>>>> Hi Sajith,
>>>>
>>>> Could you please provide some details about how you are actually doing
>>>> these performance tests. For example, what is deployment model?  How are
>>>> you generating these workloads/events?
>>>>
>>>> Thanks
>>>>
>>>> Malith
>>>>
>>>> On Thu, Jun 23, 2016 at 11:28 AM, Sajith Ravindra <[email protected]>
>>>> wrote:
>>>>
>>>>> Hi all,
>>>>>
>>>>> This is to give an update on the performance study we conducted on
>>>>> is-analytics server on last. The idea of this test round was to evaluate
>>>>> the performance of Siddhi queries used for is-analytics, therefore we
>>>>> disabled event stream persistence and spark for this test.
>>>>>
>>>>> This test was conducted on a standalone DAS server with Xms2g and
>>>>> Xmx4g.
>>>>>
>>>>> On the initial round when input TPS reaches ~20K, the server went OOM
>>>>> after consuming around 1M events. The reason for this was the events
>>>>> accumulated inside 7 1min time batch windows used inside.  To overcome 
>>>>> this
>>>>> we implemented an extension to siddhi which allows us to avoid duplicating
>>>>> the window.
>>>>>
>>>>> After removing duplicate windows the server was able to consume events
>>>>> at a rate of ~22K, but there were fluctuations (see the graph bellow) of
>>>>> the throughput. With analysis, we found that intense GC causes this. We
>>>>> suspect that this intense GC  is caused when expiring a large number of
>>>>> events accumulated inside 1-minute window. To  overcome this we are
>>>>> planning to batch events in 1-second windows and then accumulate 1second
>>>>> batches in 1 min window in order to stop accumulating a large number of
>>>>> events.
>>>>>
>>>>>
>>>>> As the next steps, we are planning to test the performance with event
>>>>> stream persistence and then move on to check the performance in DAS 
>>>>> minimum
>>>>> HA mode.
>>>>>
>>>>> We will keep updating this thread with our findings.
>>>>>
>>>>> Please share your thought, suggestions on this.
>>>>>
>>>>> Thanks
>>>>> *,Sajith Ravindra*
>>>>> Senior Software Engineer
>>>>> WSO2 Inc.; http://wso2.com
>>>>> lean.enterprise.middleware
>>>>>
>>>>> mobile: +94 77 2273550
>>>>> blog: http://sajithr.blogspot.com/
>>>>> <http://lk.linkedin.com/pub/shani-ranasinghe/34/111/ab>
>>>>>
>>>>> _______________________________________________
>>>>> Architecture mailing list
>>>>> [email protected]
>>>>> https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture
>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> Malith Jayasinghe
>>>>
>>>>
>>>> WSO2, Inc. (http://wso2.com)
>>>> Email   : [email protected]
>>>> Mobile : 0770704040
>>>> Lean . Enterprise . Middleware
>>>>
>>>> _______________________________________________
>>>> Architecture mailing list
>>>> [email protected]
>>>> https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture
>>>>
>>>>
>>>
>>> _______________________________________________
>>> Architecture mailing list
>>> [email protected]
>>> https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture
>>>
>>>
>>
>>
>> --
>> Malith Jayasinghe
>>
>>
>> WSO2, Inc. (http://wso2.com)
>> Email   : [email protected]
>> Mobile : 0770704040
>> Lean . Enterprise . Middleware
>>
>> _______________________________________________
>> Architecture mailing list
>> [email protected]
>> https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture
>>
>>
>
> _______________________________________________
> Architecture mailing list
> [email protected]
> https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture
>
>
_______________________________________________
Architecture mailing list
[email protected]
https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture

Reply via email to