289

I would like to avoid doing this by launching the process from a monitoring app.

11 Answers 11

374

On Linux with the ps from procps(-ng) (and most other systems since this is specified by POSIX):

ps -o etime= -p "$$" 

Where $$ is the PID of the process you want to check. This will return the elapsed time in the format [[dd-]hh:]mm:ss.

Using -o etime tells ps that you just want the elapsed time field, and the = at the end of that suppresses the header (without, you get a line which says ELAPSED and then the time on the next line; with, you get just one line with the time).

Or, with newer versions of the procps-ng tool suite (3.3.0 or above) on Linux or on FreeBSD 9.0 or above (and possibly others), use:

ps -o etimes= -p "$$"

(with an added s) to get time formatted just as seconds, which is more useful in scripts.

On Linux, the ps program gets this from /proc/$$/stat, where one of the fields (see man proc) is process start time. This is, unfortunately, specified to be the time in jiffies (an arbitrary time counter used in the Linux kernel) since the system boot. So you have to determine the time at which the system booted (from /proc/stat), the number of jiffies per second on this system, and then do the math to get the elapsed time in a useful format.

It turns out to be ridiculously complicated to find the value of HZ (that is, jiffies per second). From comments in sysinfo.c in the procps package, one can A) include the kernel header file and recompile if a different kernel is used, B) use the posix sysconf() function, which, sadly, uses a hard-coded value compiled into the C library, or C) ask the kernel, but there's no official interface to doing that. So, the ps code includes a series of kludges by which it determines the correct value. Wow.

So it's convenient that ps does that all for you. :)

As user @336_ notes, on Linux (this is not portable), you can use the stat command to look at the access, modification, or status change dates for the directory /proc/$$ (where again $$ is the process of interest). All three numbers should be the same, so

stat -c%X /proc/$$

will give you the time that process $$ started, in seconds since the epoch. That still isn't quite what you want, since you still need to do the math to subtract that from the current time to get elapsed time — I guess something like date +%s --date="now - $( stat -c%X /proc/$$ ) seconds" would work, but it's a bit ungainly. One possible advantage is that if you use the long-format output like -c%x instead of -c%X, you get greater resolution than whole-number seconds. But, if you need that, you should probably use process-auditing approach because the timing of running the stat command is going to interfere with accuracy.

16
  • 1
    Hi! Is etime= a typo? I can only find etime in the man pages.
    – Kent Pawar
    Jun 27, 2013 at 13:01
  • 17
    @KentPawar It's not a typo. The empty = suppresses the header. Try it without, or try ps -p $$ -o etime="Silly Header Here"
    – mattdm
    Jun 27, 2013 at 13:28
  • 4
    ps -p $(pgrep find) -o etime=
    – mafrosis
    Sep 11, 2013 at 23:07
  • 1
    Nice. I prefer etimes myself as then it's machine readable Jul 15, 2015 at 9:35
  • 1
    @alexmurray That just calls sysconf() and therefore gives you the hard-coded value from the C library, as noted, doesn't it?
    – mattdm
    Apr 1, 2016 at 1:55
43

Portable:

% ps -o stime,time $$
STIME     TIME
Jan30 00:00:06

i.e. that shell was started on January 30 and totaled about 6 seconds of CPU time.

There may be more precise or more parseable but less portable ways to get this information. Check the documentation of your ps command or your proc filesystem.

Under Linux, this information lives in /proc/$pid/stat.

awk '{print "CPU time: " $14+$15; print "start time: " $22}' /proc/$$/stat

The CPU time is in jiffies; I don't know offhand how to find the jiffy value from the shell. The start time is relative to the boot time (found in /proc/uptime).

3
  • 3
    Finding the value of HZ (that is, jiffies per second) turns out to be ridiculously complicated! From comments in the sysinfo.c in the procps package, one can a) include the kernel header file (and recompile if a different kernel is used, b) use the posix sysconf() function, which, sadly, uses a hard-coded value compiled into the c library, or c) ask the kernel, and there's no official interface to doing that. So, the code includes a series of kludges by which it determines the correct value. Wow.
    – mattdm
    Feb 22, 2011 at 21:54
  • 1
    The ps manpage states that time is "cumulative CPU time". I think what the OP was looking for is etime, or "the elapsed time since the process was started". pubs.opengroup.org/onlinepubs/000095399/utilities/ps.html
    – rinogo
    Apr 23, 2014 at 20:35
  • 1
    Not so "portable" after all: "ps: stime: keyword not found" on FreeBSD. It does at least support etime, though.
    – n.st
    May 23, 2016 at 11:34
24
ps -eo pid,comm,cmd,start,etime | grep -i X

X is the name of the process

2
  • 2
    should probably add a grep -v grep. Jun 30, 2014 at 20:07
  • ps -o pid,comm,cmd,start,etime -p X to look at PID X. Jul 9, 2018 at 17:06
14

ps takes a -o option to specify the output format, and one of the available columns is etime. According to the man page:

etime - elapsed time since the process was started, in the form [[dd-]hh:]mm:ss.

Thus you can run this to get the PID and elapsed time of every process:

$ ps -eo pid,etime

If you want the elapsed time of a particular PID (e.g. 12345), you can do something like:

$ ps -eo pid,etime | awk '/^12345/ {print $2}'

(Edit: Turns out there's a shorter syntax for the above command; see mattdm's answer)

9

Unsure why this has not yet been suggested: on Linux you can stat() the /proc/[nnn] directory for your PID.

This behavior is explicitly designed to return the process start time, which it can do at high resolution, and which the kernel can do accurately without the jiffies hacks since the kernel can (obviously) simply check the relevant information. The access, data-modification and status change fields all return the process start time.

Best of all, you can use stat(1) at the shell, or the appropriate binding to stat(2) from $favorite_programming_language, so you may not even need to launch an external process.

NOTE that this does not work with /usr/compat/linux/proc on FreeBSD; the access/modification/status-change times returned are the current time, and the birth time is the UNIX epoch. Quite stupid the support isn't there if you ask me.

3
  • Where in the output of stat do I see the info? I only see Access, Modify, and Change.
    – tshepang
    Jan 8, 2017 at 22:52
  • @Tshepang Note that those values are all the same, and they are the process start time. You still have to do the math, but this is definitely better than trying to figure out jiffies as noted in my answer.
    – mattdm
    Mar 2, 2017 at 16:53
  • You call it like this: stat /proc/4480 This will give you the Birth, Change, Modify and Access dates of the process. If you need the process id, just use "top"
    – user890332
    Aug 21, 2019 at 17:05
2

If you can run time and then execute a command you will get exactly what you are looking for. You cannot do this against an already-running command.

[0] % time sleep 20

sleep 20 0.00s user 0.00s system 0% cpu 20.014 total

1
  • Do you know how can I do it on a running process monitoring until it ends?
    – lrkwz
    Nov 21, 2012 at 22:56
2

$ ps -eo lstart get start time

$ ps -eo etime get duration/elapsed time

$ ps -eo pid,lstart,etime | grep 61819
  PID                   STARTED     ELAPSED
  61819 Mon Sep 17 03:01:35 2018    07:52:15

61819 is the process id.

1
1

you can get the start time of the process by looking at the stat of the stat file produced by proc, format it using date and subtract it from the current time:

echo $(( $(date +%s) - $(date -d "$(stat /proc/13494/stat | grep Modify | sed 's/Modify: //')" +%s) ))

where 13494 is your process' pid

1

Time elapsed in seconds: expr $(date +"%s") - $(stat -c%X /proc/<PID HERE>)

2
  • This seems to me to be a very slight variation of one that mattdm mentioned already: date +%s --date="now - $( stat -c%X /proc/$$
    – Jeff Schaller
    Mar 28, 2019 at 17:17
  • That one didn't work for me on my very minimal Alpine docker instance so I wrote this one
    – Shardj
    Mar 28, 2019 at 17:31
0

A handy function based on "ps -o etime=" and "bc" to help with the maths and the zero left padded numbers.

Give it the pid and get the running minutes back.

function running_time (){
  # correct up to 100 days! #08:28:40 #53:32 #15-01:23:00
  p=$1 ; i=$(ps -o etime= $p) ; i=$(echo $i) ;

  len=${#i}
  [[ $len == 5  ]] && i="00-00:$i" ; [[ $len == 8  ]] && i="00-$i"
  [[ $len == 10 ]] && i="0$i"

  mins=$(echo ${i:0:2}*24*60+${i:3:2}*60+${i:6:2}|bc)
  echo $mins
}
0

I wanted it in seconds and I came up with (pure bash):

runtime=$(ps -o etime= -p <pid>)
runtime=$(date -d "1970-01-01 $((10#${runtime: -8: 2})):$((10#${runtime: -5: 2})):$((10#${runtime: -2: 2}))Z + $((10#$(echo "$runtime" | grep -oP "[0-9]+(?=-)"))) days" +%s)

The problem with etime is, that some values are optional.

Because of that I used $((10#string)) to convert empty strings to 0 and by that the final date string will be something like this for a short running process:

date -d "1970-01-01 0:0:7Z + 0 days" +%s

and like this for a long runnning processs:

date -d "1970-01-01 12:45:11 + 253 days" +%s

Using date to convert to seconds was inspired by this answer.

Not the answer you're looking for? Browse other questions tagged or ask your own question.