Several papers, so far, on Oracle RAC have presented further evidence of poor scalability for transactional workloads (OLTP). This follows similar results reported at Hotsos 2007. This is quite shocking, given the 10 or so years Oracle Corp. has spent on trying to improve the scalability of OPS.
I was feeling a little defensive about presenting my talk on PerfViz (performance visualization) on Wednesday, but at least one paper by Tanel Poder had some interesting scripted demos which I now believe provides a nice segue for my own talk on better visualization techniques for Oracle performance data. One of his drill-down demos showed the conflicted situation where the Oracle Wait interface reported that the CPU was busy, while UNIX performance tools showed %cpu was zero. A stack trace showed that the UNIX process was suspended waiting on file I/O, but the Oracle tool did know that state information.
The last presentation I attended was by Gerwin Hendriksen who showed how Oracle's data mining tools (an add-on package) could be used to do both variable reduction (factorial ANOVAR) and regression analysis on system-wide performance metrics to detect bottlenecks that contributed to response time variance. His approach is very similar to the technique I use in Chapter 8 of my Guerrilla Capacity Planning book. Factorial experiment techniques are presented in the Guerrilla Data Analysis Techniques class.
On the non-tech side. This is March ... in Dallas ... Dallas, Texas. Today, looking out of my hotel room window, it's sunny and about 70*F (21*C). Last night, I was invited by the AMIS contingent from Holland to join them for dinner at a steak house across the street from the hotel. It was raining when we went, but by the time we were heading back to the hotel, it was snowing!
No comments:
Post a Comment