一、问题提出
会议中有同学提到使用mktime遇到一些问题： 1) 设置tm_isdst后速度很慢 2) 设置TZ环境变量提速极大 所以想调查下具体情况。
mktime真的这么慢？如果是，为什么？
二、测试和检验
环境（不同环境可能结果迥异，以下所述仅对本环境有效）
$ cat /proc/version Linux version --tlinux2-.tl2 (mockbuild@TENCENT64.site) (gcc version (Red Hat -) (GCC) ) # SMP Fri Apr :: CST $ getconf -a | grep glibc -i GNU_LIBC_VERSION glibc 2.17
首先写了个简单的mktime测试。
#include <sys/time.h>
#include <stdlib.h>
#include <stdio.h>
#include <time.h>
typedef int64_t timestamp_t;
static timestamp_t get_timestamp()
{
struct timeval tv = {};
gettimeofday(&tv, );
+ (timestamp_t)tv.tv_usec;
}
static void call_mktime(int isdst)
{
struct tm tm = {};
tm.tm_year = - ;
tm.tm_mon = - ;
tm.tm_mday = ;
tm.tm_hour = ;
tm.tm_min = ;
tm.tm_sec = ;
tm.tm_isdst = isdst;
== mktime(&tm)) {
abort();
}
}
int main()
{
, };
for (const auto &isdst: isdsts) {
timestamp_t t1 = get_timestamp();
;
; i < N; ++i) {
call_mktime(isdst);
}
timestamp_t t2 = get_timestamp();
printf("isdst=%d rounds %d avg cost %4.2f us\n", isdst, N, 1.0*(t2-t1)/N);
}
;
}
跑一下，得到结果如下：
$ TZ="Asia/Shanghai" ./app_test_mktime
isdst=1 rounds 100 avg cost 484.32 us
isdst=0 rounds 100 avg cost 2.17 us
还真的很慢啊！慢的掉渣了！
But！真的是这样吗？
要不试试其他时区？就用美国东部时间好了。
$ TZ="US/Eastern" ./app_test_mktime
isdst=1 rounds 100 avg cost 2.31 us
isdst=0 rounds 100 avg cost 0.20 us
奇迹发生了，不慢啊，也就几微秒而已，虽然慢了一点点，但绝对没有那么夸张。
三、基本结论
跟着这个思路，稍微扩大一下可变的参数，包括下面几个因子：
- 日历时间
- 时区配置
- 输入isdst
最后跑出一个结果（源代码见后）：
|
|
|
|
isdst |
||
|
日历时间 |
时区配置 |
夏令时 |
1 |
0 |
-1 |
|
1688-06-01 02:00 |
Asia/Shanghai |
N |
103.95 |
0.23 |
0.23 |
|
|
US/Eastern |
N |
114.6 |
0.26 |
0.23 |
|
|
America/Jujuy |
N |
125.26 |
0.27 |
0.25 |
|
1960-06-01 02:00 |
Asia/Shanghai |
N |
158.9 |
0.3 |
0.29 |
|
|
US/Eastern |
Y |
0.6 |
4.3 |
0.39 |
|
|
America/Jujuy |
Y |
0.48 |
67.24 |
0.34 |
|
1986-06-01 02:00 |
Asia/Shanghai |
Y |
0.48 |
2.47 |
0.3 |
|
|
US/Eastern |
Y |
0.44 |
2.73 |
0.3 |
|
|
America/Jujuy |
N |
54.67 |
0.34 |
0.32 |
|
2016-01-01 02:00 |
Asia/Shanghai |
N |
501.6 |
0.76 |
0.76 |
|
|
US/Eastern |
N |
4.49 |
0.32 |
0.31 |
|
|
America/Jujuy |
N |
507.45 |
0.81 |
0.78 |
|
2016-05-01 02:00 |
Asia/Shanghai |
N |
505.49 |
0.7 |
0.7 |
|
|
US/Eastern |
Y |
0.64 |
3.77 |
0.31 |
|
|
America/Jujuy |
N |
514.04 |
0.76 |
0.77 |
最后两列是一次mktime调用消耗的微秒数。 注意America/Jujuy这个时区，非常有趣。
从这张表格可以总结出一个基本结论： 当tm_isdst设置不当时，调用mktime会消耗更多的时间。
- 如果当时的日历时间是夏令时，那么
isdst=1速度比isdst=0快 ； - 如果当时的日立时间是常规时，那么
isdst=1速度比isdst=0慢 ； - 调用mktime，可以传入
isdst=-1，让glibc根据时区自动决定DST标记。
The mktime() function converts a broken-down time structure, expressed as local time, to calendar time representation. The function ignores the values supplied by the caller in the tm_wday and tm_yday fields. The value specified in the tm_isdst field informs mktime() whether or not daylight saving time (DST) is in effect for the time supplied in the tm structure: a positive value means DST is in effect; zero means that DST is not in effect; and a negative value means that mktime() should (use timezone information and system databases to) attempt to determine whether DST is in effect at the specified time.
至于KM文章中提到的设置TZ环境变量导致的性能差异，是非常小的，有兴趣的可以做个测试。
四、进一步调查
为什么isdst配置不当时，速度会相差这么多？这和mktime的实现有关。
mktime转换年月日格式的时间到时间戳，分几个步骤
- 6次循环，猜测得到一个时间戳t1，调用localtime(t1)能够得到正确的年月日表示。但它的夏令时标记（isdst）可能和用户传递进来的不一致。
- 如果夏令时标记和用户调用传递的isdst不同(0 vs 1, 1 vs 0)，以t1为基准，前后搜寻合适的日历时，如果找到一个日历时，它的DST标记符合，以该日历时所处的DST为准。如果无法搜寻到合适的结果，直接返回t1
- 搜寻思想：以当前时间为中心，前后搜寻，找到一个与输入参数匹配的夏令时/冬令时区间，以该区间的配置来校准t1。
- 搜寻算法：
- 步长：601200秒，这是所有夏令时区间中最短的一个周期：7天，时区为：America/Recife
- 范围：以t1为中心的536454000秒区间。这是所有夏令时区间中最长的一个周期：17年，时区为America/Jujuy。范围：[t-536454000/2+601200,t+536454000/2+601200]，最大故迭代次数894次。
- 迭代： 对当前时间戳tx调用localtime(tx)，若结果的isdst和输入的isdst相同，命中，跳出循环。否则继续。
- 命中：根据命中时的DST设置，找到正确的时间戳t3，转换成功。
迭代次数越多，耗时就越多。如果步骤1转换的posix time距离最近的匹配区间（夏令时/冬令时）很远，搜寻耗时就很长。
Asio/Shanghai的夏令时从1991年废除，而US/Eastern每年都有夏令时，所以，大部分情况下前者的迭代次数远大于后者，这也能很好的解释上面的图表。
用zdump看一下不同时区数据库的信息，注意1688和1986两个测试日历时间的选取。
$ zdump -v Asia/Shanghai Asia/Shanghai -9223372036854775808 = NULL Asia/Shanghai -9223372036854689408 = NULL Asia/Shanghai Mon Dec 31 15:54:16 1900 UTC = Mon Dec 31 23:59:59 1900 LMT isdst=0 gmtoff=29143 Asia/Shanghai Mon Dec 31 15:54:17 1900 UTC = Mon Dec 31 23:54:17 1900 CST isdst=0 gmtoff=28800 Asia/Shanghai Sun Jun 2 15:59:59 1940 UTC = Sun Jun 2 23:59:59 1940 CST isdst=0 gmtoff=28800 Asia/Shanghai Sun Jun 2 16:00:00 1940 UTC = Mon Jun 3 01:00:00 1940 CDT isdst=1 gmtoff=32400 Asia/Shanghai Mon Sep 30 14:59:59 1940 UTC = Mon Sep 30 23:59:59 1940 CDT isdst=1 gmtoff=32400 Asia/Shanghai Mon Sep 30 15:00:00 1940 UTC = Mon Sep 30 23:00:00 1940 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Mar 15 15:59:59 1941 UTC = Sat Mar 15 23:59:59 1941 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Mar 15 16:00:00 1941 UTC = Sun Mar 16 01:00:00 1941 CDT isdst=1 gmtoff=32400 Asia/Shanghai Tue Sep 30 14:59:59 1941 UTC = Tue Sep 30 23:59:59 1941 CDT isdst=1 gmtoff=32400 Asia/Shanghai Tue Sep 30 15:00:00 1941 UTC = Tue Sep 30 23:00:00 1941 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat May 3 15:59:59 1986 UTC = Sat May 3 23:59:59 1986 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat May 3 16:00:00 1986 UTC = Sun May 4 01:00:00 1986 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 13 14:59:59 1986 UTC = Sat Sep 13 23:59:59 1986 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 13 15:00:00 1986 UTC = Sat Sep 13 23:00:00 1986 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 11 15:59:59 1987 UTC = Sat Apr 11 23:59:59 1987 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 11 16:00:00 1987 UTC = Sun Apr 12 01:00:00 1987 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 12 14:59:59 1987 UTC = Sat Sep 12 23:59:59 1987 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 12 15:00:00 1987 UTC = Sat Sep 12 23:00:00 1987 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 9 15:59:59 1988 UTC = Sat Apr 9 23:59:59 1988 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 9 16:00:00 1988 UTC = Sun Apr 10 01:00:00 1988 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 10 14:59:59 1988 UTC = Sat Sep 10 23:59:59 1988 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 10 15:00:00 1988 UTC = Sat Sep 10 23:00:00 1988 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 15 15:59:59 1989 UTC = Sat Apr 15 23:59:59 1989 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 15 16:00:00 1989 UTC = Sun Apr 16 01:00:00 1989 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 16 14:59:59 1989 UTC = Sat Sep 16 23:59:59 1989 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 16 15:00:00 1989 UTC = Sat Sep 16 23:00:00 1989 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 14 15:59:59 1990 UTC = Sat Apr 14 23:59:59 1990 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 14 16:00:00 1990 UTC = Sun Apr 15 01:00:00 1990 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 15 14:59:59 1990 UTC = Sat Sep 15 23:59:59 1990 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 15 15:00:00 1990 UTC = Sat Sep 15 23:00:00 1990 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 13 15:59:59 1991 UTC = Sat Apr 13 23:59:59 1991 CST isdst=0 gmtoff=28800 Asia/Shanghai Sat Apr 13 16:00:00 1991 UTC = Sun Apr 14 01:00:00 1991 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 14 14:59:59 1991 UTC = Sat Sep 14 23:59:59 1991 CDT isdst=1 gmtoff=32400 Asia/Shanghai Sat Sep 14 15:00:00 1991 UTC = Sat Sep 14 23:00:00 1991 CST isdst=0 gmtoff=28800 Asia/Shanghai 9223372036854689407 = NULL Asia/Shanghai 9223372036854775807 = NULL $ zdump -v US/Eastern US/Eastern -9223372036854775808 = NULL US/Eastern -9223372036854689408 = NULL US/Eastern Sun Nov 18 16:59:59 1883 UTC = Sun Nov 18 12:03:57 1883 LMT isdst=0 gmtoff=-17762 US/Eastern Sun Nov 18 17:00:00 1883 UTC = Sun Nov 18 12:00:00 1883 EST isdst=0 gmtoff=-18000 US/Eastern Sun Mar 31 06:59:59 1918 UTC = Sun Mar 31 01:59:59 1918 EST isdst=0 gmtoff=-18000 US/Eastern Sun Mar 31 07:00:00 1918 UTC = Sun Mar 31 03:00:00 1918 EDT isdst=1 gmtoff=-14400 US/Eastern Sun Oct 27 05:59:59 1918 UTC = Sun Oct 27 01:59:59 1918 EDT isdst=1 gmtoff=-14400 US/Eastern Sun Oct 27 06:00:00 1918 UTC = Sun Oct 27 01:00:00 1918 EST isdst=0 gmtoff=-18000 US/Eastern Sun Mar 30 06:59:59 1919 UTC = Sun Mar 30 01:59:59 1919 EST isdst=0 gmtoff=-18000 US/Eastern Sun Mar 30 07:00:00 1919 UTC = Sun Mar 30 03:00:00 1919 EDT isdst=1 gmtoff=-14400 …省略若干… US/Eastern Sun Mar 9 06:59:59 2498 UTC = Sun Mar 9 01:59:59 2498 EST isdst=0 gmtoff=-18000 US/Eastern Sun Mar 9 07:00:00 2498 UTC = Sun Mar 9 03:00:00 2498 EDT isdst=1 gmtoff=-14400 US/Eastern Sun Nov 2 05:59:59 2498 UTC = Sun Nov 2 01:59:59 2498 EDT isdst=1 gmtoff=-14400 US/Eastern Sun Nov 2 06:00:00 2498 UTC = Sun Nov 2 01:00:00 2498 EST isdst=0 gmtoff=-18000 US/Eastern Sun Mar 8 06:59:59 2499 UTC = Sun Mar 8 01:59:59 2499 EST isdst=0 gmtoff=-18000 US/Eastern Sun Mar 8 07:00:00 2499 UTC = Sun Mar 8 03:00:00 2499 EDT isdst=1 gmtoff=-14400 US/Eastern Sun Nov 1 05:59:59 2499 UTC = Sun Nov 1 01:59:59 2499 EDT isdst=1 gmtoff=-14400 US/Eastern Sun Nov 1 06:00:00 2499 UTC = Sun Nov 1 01:00:00 2499 EST isdst=0 gmtoff=-18000 US/Eastern 9223372036854689407 = NULL US/Eastern 9223372036854775807 = NULL
五、附录
1. test_timezone.cpp
测试不同时区的mktime性能
#include <sys/time.h>
#include <stdlib.h>
#include <stdio.h>
#include <time.h>
typedef int64_t timestamp_t;
static timestamp_t get_timestamp()
{
struct timeval tv = {};
gettimeofday(&tv, );
+ (timestamp_t)tv.tv_usec;
}
struct calendar_time {
int year;
int month;
int day;
int hour;
int minute;
int second;
};
static void call_mktime(
const calendar_time &calendar,
int isdst)
{
struct tm tm = {};
tm.tm_year = calendar.year - ;
tm.tm_mon = calendar.month- ;
tm.tm_mday = calendar.day;
tm.tm_hour = calendar.hour;
tm.tm_min = calendar.minute;
tm.tm_sec = calendar.second;
tm.tm_isdst = isdst;
== mktime(&tm)) {
abort();
}
}
int main()
{
const char *timeonzes[] = { "Asia/Shanghai", "US/Eastern", "America/Jujuy"};
calendar_time times[] = {
{, , , , , },
{, , , , , },
{, , , , , },
{, , , , , },
{, , , , , },
{, , , , , },
{, , , , , },
};
, , -};
for (const auto &calendar: times) {
];
for (const auto &tz: timeonzes) {
for (const auto &isdst: isdsts) {
setenv();
timestamp_t t1 = get_timestamp();
;
; i < N; ++i) {
call_mktime(calendar, isdst);
}
timestamp_t t2 = get_timestamp();
printf("calendar: %04d-%02d-%02d %02d:%02d:%02d ",
calendar.year, calendar.month, calendar.day,
calendar.hour, calendar.minute, calendar.second);
printf("%-20s isdst=%2d rounds %d avg cost %4.2f us\n", tz, isdst, N, 1.0*(t2-t1)/N);
}
printf("\n");
}
printf("-------------------------------------------\n");
}
;
}
2. test_setenv.cpp
测试setenv("TZ")和不设置时的性能差别。
#include <sys/time.h>
#include <stdlib.h>
#include <stdio.h>
#include <time.h>
typedef int64_t timestamp_t;
static timestamp_t get_timestamp()
{
struct timeval tv = {};
gettimeofday(&tv, );
+ (timestamp_t)tv.tv_usec;
}
struct calendar_time {
int year;
int month;
int day;
int hour;
int minute;
int second;
};
static void call_mktime(
const calendar_time &calendar,
int isdst)
{
struct tm tm = {};
tm.tm_year = calendar.year - ;
tm.tm_mon = calendar.month- ;
tm.tm_mday = calendar.day;
tm.tm_hour = calendar.hour;
tm.tm_min = calendar.minute;
tm.tm_sec = calendar.second;
tm.tm_isdst = isdst;
== mktime(&tm)) {
abort();
}
}
int main()
{
const char *tz = "Asia/Shanghai";
calendar_time times[] = {
{, , , , , },
{, , , , , },
{, , , , , },
{, , , , , },
};
for (const auto &calendar: times) {
printf("calendar: %04d-%02d-%02d %02d:%02d:%02d\n",
calendar.year, calendar.month, calendar.day,
calendar.hour, calendar.minute, calendar.second);
{
unsetenv("TZ");
timestamp_t t1 = get_timestamp();
;
; i < N; ++i) {
call_mktime(calendar, );
}
timestamp_t t2 = get_timestamp();
printf("unsetenv %-20s rounds %d avg cost %4ld us\n", tz, N, (t2-t1)/N);
}
{
setenv();
timestamp_t t1 = get_timestamp();
;
; i < N; ++i) {
call_mktime(calendar, );
}
timestamp_t t2 = get_timestamp();
printf("setenv %-20s rounds %d avg cost %4ld us\n", tz, N, (t2-t1)/N);
}
printf("\n");
}
;
}