C
ChUIMonster
Guest
Excellent baseline questions from Simon -- it really helps to know what the target is. If the customer keeps moving the goal posts it is very hard to every make them happy. When the customer really and truly wants the process to be as short as possible you have to spend time testing various options and doing dry runs. Every system is different and your results *will* vary from what someone's Powerpoint or forum posting has to say. None the less... Paul's talk from the EMEA PUG is quite good and covers a lot. I second the recommendation for binary d&l. It is almost always going to be faster. If the data is already in type 2 storage areas and speed is your main objective then a binary dump with "-index 0" will probably be fastest. You should also probably use -RO rather than starting a server. You will need to test it but, in many cases, starting one dump per (real) CPU in parallel will be fastest. Getting the scripts written to balance the threads and get them all to finish more or less together can be challenging. For the load I would use -r rather than -i, you get pretty much all the same benefits with better recoverability (neither option is suitable for *production* but they are useful when building a db from scratch). In a virtual environment there are a million or so things that can fool you. Be especially careful not to mistake vCPUs for real CPUs - a hyper threaded vCPU is not worth much trying to do things like allocate one thread per CPU. You are probably better off not counting hyper threaded CPUs at all. Your IO subsystem is going to make a huge difference too.
Continue reading...
Continue reading...