Performance: Pipelines vs Process Substitution

bashperformancepipeprocess-substitutionshell

I tend to use pipelines in my bash scripts over process substitution in most situations, especially in cases of using multiple sets of commands as it seems more readable to do ... | ... | ... over ... < <(... < <(...)).

I'm wondering though why using process substitution is much faster in some situations than using a pipeline.

To test this, I timed two scripts using 10000 iterations of the same attached commands with one using a pipeline and another using process substitution.

Scripts:

pipeline.bash:

for i in {1..10000}; do
    echo foo bar |
    while read; do
        echo $REPLY >/dev/null
    done
done

proc-sub.bash

for i in {1..10000}; do
    while read; do
        echo $REPLY >/dev/null
    done < <(echo foo bar)
done

Results:

~$ time ./pipeline.bash

real    0m17.678s
user    0m14.666s
sys     0m14.807s

~$ time ./proc-sub.bash

real    0m8.479s
user    0m4.649s
sys     0m6.358s

I know that pipelines create a sub process whereas process substitution creates a named pipe or some file in /dev/fd, but am unclear about how those differences impact performance.

Best Answer

Doing same strace, you can see the differences:

With pipe:

$ strace -c ./pipe.sh 
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 57.89    0.103005           5     20000           clone
 40.81    0.072616           2     30000     10000 wait4
  0.58    0.001037           0    120008           rt_sigprocmask
  0.40    0.000711           0     10000           pipe

With proc-sub:

$ strace -c ./procsub.sh 
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 85.08    0.045502           5     10000           clone
  3.25    0.001736           0     90329       322 read
  2.12    0.001133           0     20009           open
  2.03    0.001086           0     50001           dup2

With above statistics, you can see pipe create more child processes (clone syscall) and spending many times to wait child process (wait4 syscall) to finish for parent process to continue executing.

Process substitution is not. It can read directly from child processes. Process substitution is performed at the same time with parameter and variable expansion, the command in Process Substitution run in background. From bash manpage:

Process Substitution
       Process  substitution  is supported on systems that support named pipes
       (FIFOs) or the /dev/fd method of naming open files.  It takes the  form
       of  <(list) or >(list).  The process list is run with its input or out‐
       put connected to a FIFO or some file in /dev/fd.  The name of this file
       is  passed  as  an argument to the current command as the result of the
       expansion.  If the >(list) form is used, writing to the file will  pro‐
       vide  input  for list.  If the <(list) form is used, the file passed as
       an argument should be read to obtain the output of list.

       When available, process substitution is performed  simultaneously  with
       parameter  and variable expansion, command substitution, and arithmetic
       expansion.

Update

Doing strace with statistics from child processes:

With pipe:

$ strace -fqc ./pipe.sh 
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 70.76    0.215739           7     30000     10000 wait4
 28.04    0.085490           4     20000           clone
  0.78    0.002374           0    220008           rt_sigprocmask
  0.17    0.000516           0    110009     20000 close
  0.15    0.000456           0     10000           pipe

With proc-sub:

$ strace -fqc ./procsub.sh 
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 52.38    0.033977           3     10000           clone
 32.24    0.020913           0     96070      6063 read
  5.24    0.003398           0     20009           open
  2.34    0.001521           0    110003     10001 fcntl
  1.87    0.001210           0    100009           close
Related Question