четверг, 8 мая 2008 г.


Below are a list of traps, and some wisdom which may help you find a faster, or more accurate diagnosis.
• Don't confuse the symptom with the problem
(e.g. latch free wait event is a symptom, not the problem)
• No statistic is an island i.e. don’t rely on a single piece of evidence in isolation to make a diagnosis
(e.g. the buffer cache hit ratio can often be misleading; similarly with other rolled-up statistics)
• Don't jump to conclusions
• Don’t be sidetracked by irrelevant statistics (there are lots of them)
(out of the 255 V$SYSSTAT statistics, there are approximately 15 being useful for 99% of issues)
• If it isn’t broken, think twice before fixing it (or don’t fix it at all)
• Don’t be predisposed to finding problems you know the solutions to - this is known as the old favourite.
(i.e. The problem identified and evidence sought is related to a favourite issue encountered previously, for which there is a well known solution, rather than looking for the bottleneck. Usually, the required evidence will be found to support the preconception)
• Be wary of solutions that involve fixing a serious performance problem by applying a single, simple change.
(This usually involves setting an init.ora parameter, with _underscore parameters being popular. This type of solution is known as a silver bullet)
• Make changes to a system only after you are certain of the cause of the bottleneck
(If you do make changes hastily, in the worst case performance will degrade)
• Many times, modifying the application results in significantly larger and longer term performance gains, when compared to solutions based on tweaking init.ora parameters7 (rejection of this actuality is often accompanied by the search for a silver bullet solution)
• Removing one bottleneck may result in the dynamics of the instance changing (a good reason not to implement multiple changes at once), and hence the next bottleneck to be solved may not be the current second in the list

Комментариев нет: