Years ago in the begin of this century I was working for a customer and at the end of each week the technical team was sending an email with some graphs to their management, illustrating how the performance had been that week. I was also one of the “lucky” people getting this email. The email contained response time graphs from 4 Key performance indicator processes from the business (Load Runner, every 15 minutes) and the CPU graphs of the two database (stretched cluster RAC) and application servers of their E-Business system. Every week I was just looking at the email and it made me wonder what the different graphs actually were saying, and I came to the very unpleasant conclusion….. “Almost Nothing”. So spikes at the cpu’s of the response times of the measured business processes did not have corresponding spikes in the CPU graphs, well almost never. I just started to wonder why this email was sent in the first place and if it was even possible to have meaningful information from this kind of emails. After some time I concluded that it must be possible to say something about end user response time by using data of the chain, meaning I/O, CPU, memory, statspack, etc (see below picture of some first representations):
My first attempts were based on excel and on self created linear function and models. I didn’t give it a name yet but we can say “Gerwin’s Attempt to Performance Profile” the end user process (of course with a wink this would say “GAPP”). After looking at the amount of work it took to create the models by hand in excel I thought of a more automated way to do it and came up with the idea of using data mining. In 2008 my idea was so far that I got the opportunity to present my idea at the HOTSOS symposium 2008. Also driven by a statement of Cary Millsap saying “If it doesn’t have a name it doesn’t exists”, I baptised my idea and now method “method-GAPP”.
Although the version of method-GAPP introduced in 2008 and presented in 2009 was already powerful, I continued enhancing my method to be able to have better predictions and have the service time per component also able to be determined within certain error. These last enhancements and their big impact and use have been presented at the HOTSOS symposium 2011. Currently I am working on other models and the use of the open source software R. At the UKOUG 2011 I presented the first time on using method-GAPP for mining AWR data. In the next presentation new customer cases will be presented and the use of R will be further shown.