libfast-mktime

This is a micro-optimization for mktime(3) that will speed it up around 10x to 86x in the best case and have a slight impact in the worse.

Motivation

The system mktime(3) handles time zones, tzset(), leap years, input normalization, plus the obvious date calculations, all of which can take some cycles. If your use case entails many mktime() calls within the same calendar day, there is an easy optimization which can show substantial savings.

A common use case would be parsing a large log file with many timestamps in the same day and you would like to store the timestamps in epoch time.

Caveats

1. This hack is not thread safe.
2. It does not check if the caller has switched timezones using tzset.

Algorithm

The first call, call system mktime(), memoizing the date and the beginning of that date in epoch time. Then for subsequent calls, if they are on the same date, add the requested H:M:S to that saved midnight, skipping all the date calculations. If the date is different, use the system mktime() again and memoize it.

This package provides a libfast-mktime.so.1, defining only mktime() and intended to be interpose system mktime() as above, using LD_PRELOAD.

Platforms Tested

Please let me know if something else works.

Platform	GCC	Benchmark
i686 GNU/Linux	4.6.3	~86x
x86_64 GNU/Linux	4.6.3	~86x
sun4v Solaris sparc VM	4.0.2	~10x

Building

$ make
$ make test
$ make bench                                

Usage

$ LD_PRELOAD=./libfast-mktime.so.1 /path/to/your/program

Benchmark

## Benchmarking once with stock mktime and once with wrapper.
## 3 trials of 1e6 sequential dates will be run.

./benchmark 3 1000000
  0: 0.868179 s
  1: 0.865169 s
  2: 0.866598 s
avg: 0.866649 s

LD_PRELOAD=./libfast-mktime.so.1 ./benchmark 3 1000000
  0: 0.014434 s
  1: 0.014622 s
  2: 0.015013 s
avg: 0.014690 s

License

Copyright (C) 2014 Mitchell Perilstein. Licensed under GNU LGPL Version 3. See LICENSING file for details.