Re: Queueing Theory in Oracle

  • From: Ls Cheng <exriscer@xxxxxxxxx>
  • To: Oracle Mailinglist <oracle-l@xxxxxxxxxxxxx>
  • Date: Tue, 11 Mar 2014 21:52:46 +0100

But the transaction rate I got was from v$sysmetric, that does
differentiate the 5 different application transactions




On Tue, Mar 11, 2014 at 9:51 PM, Ls Cheng <exriscer@xxxxxxxxx> wrote:

> Hi
>
> The TPC test runs 5 different type of transactions,
>
>
>    1. *New Order Transaction*
>    2. *Payment Tansaction*
>    3. *Order Status Transaction*
>    4. *Delivery Transaction*
>    5. *Stock Level Transaction*
>
>
> From the results Payment had an avg of 5ms service time, New Order 13ms,
> Delivery 15ms, Stock-Level 30ms and finally order status 261ms
>
>
>
>
>
> On Tue, Mar 11, 2014 at 9:36 PM, Jonathan Lewis <
> jonathan@xxxxxxxxxxxxxxxxxx> wrote:
>
>>
>> I was thinking more of the way the data might introduce variability in
>> the time required to execute queries - for example in an order placing
>> system a "product pick" query that supplies 20 full names and product IDs
>> for the user to choose from will take longer than a query that supplies
>> only one option.  Was there enough variation in the required service time
>> to allow a non-normal distribution ?
>>
>>
>>  Regards
>> Jonathan Lewis
>> http://jonathanlewis.wordpress.com
>> @jloracle
>>   ------------------------------
>> *From:* Ls Cheng [exriscer@xxxxxxxxx]
>> *Sent:* 11 March 2014 20:23
>> *To:* Jonathan Lewis
>> *Cc:* Oracle Mailinglist
>>
>> *Subject:* Re: Queueing Theory in Oracle
>>
>>     Hi
>>
>>  I thought the reasons of getting normal data distribution was probably
>> how the test is run. Since it's a constant 300/420 users running probably
>> 30 or 40 different SQL statements ( I dont know how many are there in a TPC
>> test), the server was only 18% loaded, the database metric I used were
>> gathered from v$sysmetrc (so I have metric rates in per second unit
>> gathered every minute), all mix together the distribution I got was normal,
>> I even took the sample data and used Cary's mdist.pl to see if the data
>> was exponentially distributed and all were rejected. After checking that
>> and think a bit then I think the normal data distribution is expected, if I
>> am running 16 TPC transactions per second and there are few in the lower
>> side a few in the higher side and most were in the middle then of course
>> it's a normal data distribution, why should I expect it to be exponentially
>> distributed?
>>
>>  I used TPS as arrival rate and little's law to get the service time
>> (used host cpu as utilization)
>>
>>  system utilization = (arrival rate * service time) / number of servers
>>
>>  The service time was normal distributed as well
>>
>>  So using the TPC test sample data, the formulas I could find (I have
>> downloaded probably 20 PPT from 7 or 8 universities statistics courses)
>> they just dont "glue" together in an Oracle Database and that is why I am
>> asking if anyone has successfully used queueing theory in Oracle so at
>> least I can get some points and see what I am dong wrong :-)
>>
>>
>> Thanks
>>
>>
>>
>>
>>
>>
>> On Tue, Mar 11, 2014 at 9:07 PM, Jonathan Lewis <
>> jonathan@xxxxxxxxxxxxxxxxxx> wrote:
>>
>>>
>>>
>>>  That's an interesting observation - but (viewed from the outside) I
>>> would be a little suspicious that the normal distribution was an artifact
>>> of the data generation mechanism and the test mechanism.
>>>
>>>
>>>
>>> Regards
>>> Jonathan Lewis
>>> http://jonathanlewis.wordpress.com
>>> @jloracle
>>>    ------------------------------
>>> *From:* Ls Cheng [exriscer@xxxxxxxxx]
>>> *Sent:* 11 March 2014 20:01
>>> *To:* Karl Arao
>>> *Cc:* Jonathan Lewis; Oracle Mailinglist
>>>
>>> *Subject:* Re: Queueing Theory in Oracle
>>>
>>>
>>>  I ran last week a couple of TPC load with 300 and 420 users then I
>>> used both transaction per second and logical reads per second metric and
>>> both showed normal data distribution and that is why I have doubts of how
>>> to use queueing theory in Oracle.
>>>
>>>  From your paper was you able to predict the change from v1 to x2
>>> without run the actual test? Then run the test and validate the prediction?
>>>
>>>
>>
>

Other related posts: