| I've always struggled getting parallel working so I played around with your example. I generated the raw list via for i in $(seq 10000); do echo "$((RANDOM % 255)).$((RANDOM % 255)).$((RANDOM % 255)).$((RANDOM % 255))" >> banlist_ipv4.raw; done Running the example, the sed '2!d' is really not working for me. It keeps throwing "bash: !d': event not found" and I can't seem to find the right escapes for it to work. I modified it to this time $(parallel --ungroup --eta --jobs 24 ipcalc --nobinary --nocolor {} :::: banlist_ipv4.raw | awk '/Network/{ print $2 }' > banlist_ipv4.formatted) Which ran in 1m3.876s I then wrote it for xargs time $(xargs --arg-file=banlist_ipv4.raw --max-procs=24 -I{} ipcalc --nobinary --nocolor {} | awk '/Network/{ print $2 }' > banlist_ipv4.formatted) which runs in 0m42.346s I'd love to understand a little bit better what you were trying to show in your example and why parallel in my example seems to be at least a third slower for the same input and task. Removing the eta calculations from the parallel example doesn't change the runtime nor does adding pv to get progress with the xargs example. It's just not a meaningful amount of cost. |
~$ parallel --citation
~$ ./test.sh
#!/bin/bash
echo "note: bench-marking a long running task"
echo "BASH_VERSION=${BASH_VERSION}"
date
#jobs=0 will spawn as many as possible at one job per core
cat "./blacklistp_p2p" | parallel --ungroup --eta --jobs 0 "ipcalc {} | sed '2!d' " | grep -Ev '^(0.|255.|127.)' >> ./blacklist_p2p_converted
date
exit 0
#note the cluster network version after test run
For your example, the only difference I saw was the ">>", which shouldn't prevent the example from running.
In practice, we saw around a 30% task completion reduction in this task due to the random queue blocking time xargs generates to preserve ordered output.
Thus, unless one pins all cpu cores with hundreds of processes for several minutes... the overall completion time may differ from what was expected.
Note, the performance likely also depends on the in-ram Linux kernel Page Cache of your filesystem, and the child process execution time variability. i.e. if you are running it in an emulated/VM/WSL environment the batching may behave differently.
On average, this toy checks/converts several hundred thousand IP "3.0.0.0 - 3.127.255.255" ranges into a CIDR subnet notation "3.0.0.0/9". My example may now contain typos as the example is quite dated.
YMMV, and I hope you are able to replicate the fun =)