How to Make Things Faster

Chapter 69. Skew

Here’s a profile for a program that took almost 22 hours to run:

	Duration
Subroutine	seconds	%	Count	Mean
single-block read	60,499	76.8%	10,013,394	0.006 042
other	18,290	23.2%	919,906	0.019 882
Total	78,789	100.0%	10,933,300	0.007 206

From this, you should be able to make some simple predictions:

1. How much time would you save if you could reduce the average “single-block read” call duration from 0.006 seconds to 0.004 seconds?

You would save 10,013,394 calls × (0.006 − 0.004) sec/call ≈ 20,000 seconds.

2. How much time would you save if you could eliminate all 10,013,394 “single-block read” calls?

You would save 60,499 seconds. Almost 17 hours.

3. How much time would you save if you could eliminate half of the “single-block read” calls?

You would save half of 60,499 seconds, right? Actually…probably not. It might work out that way, but it’s unlikely.

The snag is a data property called skew. Skew is nonuniformity in your data. A tidy little profile like the one I’ve shown here hides something important in that mean column: those ten million calls didn’t all have the same duration. Some took a lot less than 0.006 seconds, and some took a lot more. How much time you save when you eliminate half your calls depends on which calls you eliminate.

Here is a profile showing “single-block read” calls grouped by duration. It’s a histogram with 11 buckets:

	Range (seconds)		Duration
	{min ≤ duration < max}		Seconds	%	Calls
1.	0.000 000	0.000 001 ...