How to Make Things Faster

Chapter 67. Go/No-Go Predictions

Imagine that your TPS reports are slow, and that the profiles for all your TPS report executions look like this:

	Duration
Event	seconds	%	Count	Mean
disk read	1	0.1%	10	0.100
non-disk stuff	999	99.9%	4,990	0.200
Total	1,000	100.0%	5,000	0.200

Further imagine that your boss’s plan for making TPS reports run faster is to buy an expensive new super-fast storage array that will make only “disk read” calls faster. What would you say?

I’m hopeful that you would be able to say confidently that buying a storage array to improve TPS reporting would be a horrible mistake (no matter how objectionable the mean duration of 0.100 sec/call may look). If the profile is accurate, then you can be 100.000% certain that faster storage access will not improve TPS report performance by any more than 0.1%.

You can make this binary go/no-go prediction without doing any arithmetic at all. Some of the predictions you’ll make with profiles will be literally this easy. It’s what should have happened before the Payroll users in Dallas did their CPU upgrade.