/wsr/ - Worksafe Requests » Thread #708822

[32 / 2 / ?]

Teach me how to use pthread.h

Anonymous Sat 31 Aug 2019 08:30:27 No.708822 View View Reply Original Report

I'm trying to learn how to use pthread.h and there is something I don't understand.
Do I have to run pthread_join every time I create a thread with pthread_create? It seems that if I don't newly created thread doesn't activate. But if I do, the new thread starts running but the main one is suspended. Why, I thought threads are supposed to run concurrently?

Anonymous

Anonymous Sat 31 Aug 2019 11:45:03 No.708854 Report

Quoted By: >>708855 >>708868

>>708822
No, you don't call join() at all unless you want to wait for a thread to exit.

How do you know if your threads are running or not? If you're just printing stuff, are you buffering output?

Anonymous

Anonymous Sat 31 Aug 2019 11:49:38 No.708855 Report

Quoted By: >>708856 >>708868

>>708854
usually printf is flushed whenever there's a endline \n, and buffered otherwise, so make sure to add that if you aren't

Anonymous

Anonymous Sat 31 Aug 2019 11:50:41 No.708856 Report

Quoted By: >>708868

>>708855
or to force flush it:
fflush(stdout);
or to disable buffering:
setbuf(stdout, NULL);

Anonymous

Anonymous Sat 31 Aug 2019 13:10:05 No.708868 Report

Quoted By: >>708871 >>708872

>>708854
>How do you know if your threads are running or not? If you're just printing stuff, are you buffering output?
Just printing stuff, no buffering. There is no output, that's how I know.
>>708855
>>708856
I use just plain write().

Anonymous

Anonymous Sat 31 Aug 2019 13:16:33 No.708871 Report

Quoted By: >>708921

>>708868
i don't think it matters if you use write(), stdout is line-buffered by default, it may even be fully buffered in some weird systems, just try to flush it

Anonymous

Anonymous Sat 31 Aug 2019 13:17:38 No.708872 Report

Quoted By: >>708921

>>708868
post example of what you are doing

Anonymous

Anonymous Sat 31 Aug 2019 13:23:58 No.708874 Report

Quoted By: >>708921

>>708822
can't paste code on /wsr/, have you seen this? >>>/g/72543657 you don't have to join for the thread to run, paste code on /g/ of what you are trying to do or a simplified version

Anonymous

Anonymous Sat 31 Aug 2019 13:27:28 No.708875 Report

Quoted By: >>708921

what are you doing between pthread_create() and the pthread_join()? threads are "slow" to start, so if your main program is not taking much time in between, of course it will get to the join() before the thread starts running, doesn't mean the join() has anything to do with it, just the time it takes to start

Anonymous

View Same Google ImgOps iqdb SauceNAO 1563660243834.jpg, 89KiB, 701x1024

Anonymous Sat 31 Aug 2019 16:28:39 No.708921 Report

Quoted By: >>708923

>>708871
>>708872
>>708874
>>708875
I deleted everything and did what you told me to do and it works now, I don't know what was that.
Anyway it seems that the program is using way more CPU than it should (88%). I don't know whether it is related to pthread or not. Is there a tool that can tell me what is using all that CPU time? Or at least some good manual determination method?

Anonymous

Anonymous Sat 31 Aug 2019 16:34:00 No.708923 Report

Quoted By: >>708950

>>708921
use a C profiler:
http://web.archive.org/web/20141129061523/http://www.cs.utah.edu/dept/old/texinfo/as/gprof.html#SEC2

probably something more fancy than gprof exists, but that's the basic tool

Anonymous

Anonymous Sat 31 Aug 2019 18:09:08 No.708950 Report

Quoted By: >>708953 >>708957 >>708958 >>708959

>>708923
Ok, it seems that it is the function that is called by pthread that is consuming so much CPU time. But why?
Here is how the function looks like:
>the entire body of the function is in an endless while loop
>the entire body of the while loop is an if statement that tests whether 1 or more seconds passed
>if so some short code is executed, which may trigger break out from the loop
After 2-3 seconds of running gprof says that the function has been called 766 times (don't quote me on this, i closed the windows so i can't check to be sure) for some reason.

Anonymous

Anonymous Sat 31 Aug 2019 18:17:41 No.708953 Report

Quoted By: >>708962

>>708950
>I think my code does this
Why are you so averse to just posting it?

Anonymous

Anonymous Sat 31 Aug 2019 18:21:53 No.708957 Report

Quoted By: >>708958 >>708959 >>708962 >>709001

>>708950
>>the entire body of the while loop is an if statement that tests whether 1 or more seconds passed
do you have any call to "sleep()" in your infinite loop? when you have an infinite loop you should always call sleep() to free the CPU for other tasks:
https://stackoverflow.com/questions/35926364/how-important-is-it-to-call-sleep-in-a-infinite-while-loop

otherwise instructions just keep being passed to be executed by the CPU even if you are just "waiting"

Anonymous

Anonymous Sat 31 Aug 2019 18:26:03 No.708958 Report

Quoted By: >>708959 >>708962

>>708957
>>708950
sorry, link was for C++, not C, but yes, you have to call sleep() if your program is idle and you don't want it to go crazy sending instructions on an infinite loop:
https://codereview.stackexchange.com/questions/42506/using-a-for-loop-or-sleeping-to-wait-for-short-intervals-of-time/42508#42508

Anonymous

Anonymous Sat 31 Aug 2019 18:28:00 No.708959 Report

Quoted By: >>708962

>>708950
>>708957
>>708958
also if your infinite loop is just waiting for 1 second to pass and it's not doing anything else you can use sleep(1) directly rather than the loop:
https://linux.die.net/man/3/sleep

Anonymous

Anonymous Sat 31 Aug 2019 18:31:45 No.708962 Report

Quoted By: >>708967 >>708999

>>708953
Because there's no need to.
>>708957
Oh ok. I don't have sleep().
>>708958
>>708959
Got that.

Anonymous

Anonymous Sat 31 Aug 2019 18:50:39 No.708967 Report

Quoted By: >>708999 >>709000

>>708962
Bad news, sleep didn't help. I guess it's no different than what I was using: getting the current time and subtracting the previous time then checking whether the result is equal to or greater than one and if it is saving that current time as the previous value.
Also I'm not really sure whether it is the created thread because I don't think I can read gprof output properly.
No idea what it is.

Anonymous

Anonymous Sat 31 Aug 2019 19:56:35 No.708999 Report

Quoted By: >>709036

>>708967
>please take another guess at what's wrong with my program based on my account of what I think it does
>>708962
>because there's no need to

Anonymous

Anonymous Sat 31 Aug 2019 19:56:59 No.709000 Report

Quoted By: >>709036

>>708967
are you adding a number of seconds as a parameter to sleep()? like sleep(1) or sleep(0.1)?
https://linux.die.net/man/3/sleep
>I guess it's no different than what I was using: getting the current time and subtracting the previous time then checking whether the result is equal to or greater than one and if it is saving that current time as the previous value.
sleep() is different to that because it frees the CPU rather than executing constantly the comparison

Anonymous

Anonymous Sat 31 Aug 2019 19:58:53 No.709001 Report

Quoted By: >>709002 >>709011

>>708957
It's a tradeoff: if you think (on average) you're going to be doing something productive in less time than two context switches take, you should spin instead of yielding.

Anonymous

Anonymous Sat 31 Aug 2019 20:01:12 No.709002 Report

Quoted By: >>709005 >>709007

>>709001
yes, if you are doing some actual work on the infinite loop rather than just waiting for something no need to call sleep() at all

Anonymous

Anonymous Sat 31 Aug 2019 20:01:27 No.709003 Report

Quoted By: >>709036

>>708822
How many real processor cores do you actually have? If it's less than the number of threads you're creating, then you're at the mercy of the OS's scheduler as to which threads get to run, and one thread that looks important to the OS can starve the others.

This also happens on hyperthreading if you create multiple threads that all use the same functional units.

Anonymous

Anonymous Sat 31 Aug 2019 20:03:50 No.709005 Report

Quoted By: >>709008 >>709011

>>709002
Or if you're waiting on another thread, but you don't think you're going to be waiting as long as it will take to switch the core from your thread and back to it again. It's not as clear-cut as "always yield if you're waiting on something".

Anonymous

Anonymous Sat 31 Aug 2019 20:05:08 No.709007 Report

Quoted By: >>709011

>>709002
("spin" has the specific meaning of "repeatedly check if something has happened yet"; if you're doing work, you're not spinning.)

Anonymous

Anonymous Sat 31 Aug 2019 20:06:58 No.709008 Report

Quoted By: >>709012

>>709005
can't you use mutex to synchronize threads rather than sleep? or would that also cause switching of context and make it slower?

Anonymous

Anonymous Sat 31 Aug 2019 20:12:20 No.709011 Report

Quoted By:

>>709007
>>709005
>>709001
good to know, i only know the basics, have never had to use threads in anything high-performance. and i don't think i would if i could find a way to avoid it

Anonymous

Anonymous Sat 31 Aug 2019 20:16:07 No.709012 Report

Quoted By: >>709014

>>709008
On a standard OS, if you're the waiting thread, and you just tried to acquire the mutex, and were told "no", how would you proceed? Your choices are to either spin (repeatedly try to take the mutex until you succeed) or sleep (get the OS to deschedule you and wake you up when the mutex is clear).

There are costs in both cases: spinning is wasted work tying up a core, sleeping ties up the core doing useless (WRT getting stuff the user wants done done) work replacing the process using the core.

If you're a webserver receiving an upload, and the next packet could be several milliseconds away, you might want to sleep. If you're part of a game's rendering pipeline, you might just want to keep checking until the previous stage has more work for you.

Anonymous

Anonymous Sat 31 Aug 2019 20:21:34 No.709014 Report

Quoted By: >>709021

>>709012
i didn't know about that, i guess it's done with pthread_mutex_trylock()? not sure what goes on if you do a normal pthread_mutex_lock() but i guess behind the scenes it may be similar to sleep() if you can't lock the mutex immediately

Anonymous

Anonymous Sat 31 Aug 2019 20:57:36 No.709021 Report

Quoted By:

>>709014
AFAICT, POSIX doesn't specify what _lock actually does, just how it behaves. This is on purpose.

For most of POSIX's history, a home computer only had one core, so the standard allows for an OS written for a uniprocessor computer to provide a pthread implementation that relies on two threads being physically incapable of running at the same time, benefiting from simpler quicker locking primitives, no need for atomic instructions, etc. etc.

It also allows for pthreads to be implemented on an OS that does cooperative multitasking, though on such an OS you the programmer are responsible for calling _yield anywhere you might loop and never make a system call.

I would imagine _lock sleeps on most modern OSes, because that's a sensible default.

Anonymous

Anonymous Sat 31 Aug 2019 22:07:09 No.709036 Report

Quoted By: >>709041

>>708999
Ok, I'll post code if it's not what I think it is.
>>709000
>are you adding a number of seconds as a parameter to sleep()? like sleep(1) or sleep(0.1)?
Yes.
>>709003
I'm working on an old Pentium computer. So one core. Running Debian. Only two threads though. But it's certainly not that, I use many other program that make heavy use of threads and never had problems.

I think I know what it is. Since my program constantly has to do something every 1 second it certainly means that kernel grants it almost all of CPU time since there are no other programs using CPU. If I spawn another process kernel simply grants new process CPU time it needs while my program continues running normally. The system doesn't lag at all, the process doesn't lag at all, literally everything is smooth.
I'll see if I can find some docs online that confirm this.

Anonymous

Anonymous Sat 31 Aug 2019 22:39:13 No.709041 Report

Quoted By:

>>709036
>so one core
>one core
>it's like only one thread runs at a time
>But it's certainly not that,
It's that. Run it on literally anything a modern human would consider a computer and you'll see.

Capcode	All Only User Posts Only Verified Posts Only Moderator Posts Only Manager Posts Only Admin Posts Only Developer Posts Only Founder Posts
Show Posts	All Only With Images Only Without Images Only Spoiler Images Only Non-Spoiler Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

On these archives

Your latest searches

Teach me how to use pthread.h