Skip to contentSkip to conversation

Read sent/received bytes -- without opening a file in /sys or /proc

DISCUSSION Jan-Philip gehrcke 2 participants 2 messages Feb 17, 2011

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Dear list,

I'm about to write a small C program for measurements of the bandwidth on some network interface of my home router (uname -a: Linux fritz.fonwlan.box 2.6.19.2 #2 Thu Nov 18 16:35:17 CET 2010 mips GNU/Linux). I would like to accomplish very *precise* results, while keeping the absolute measurement time low.

In general, this problem requires the knowledge of time and byte differences. Therefore, two time measurements and two byte counter measurements for each, received and sent bytes, are required. Currently, I am measuring time via

clock_gettime(CLOCK_MONOTONIC, ...)

and reading the byte counters by reading files in sysfs, e.g.

     /sys/class/net/${interface}/statistics/rx_bytes.

This is very easy and works well, but reading these files four times takes some (varying) time* on my device, which introduces an error to my calculation.

Hence, I am wondering, if there is a way to receive this kind of data directly and much faster, via "kernel API". My question is related to this unanswered one from 2005:
http://linux.derkeiler.com/Mailing-Lists/RedHat/2005-02/0557.html

*On my device, reading such a file and interpreting the result as integer takes 1-10 ms, measured via:

    clock_gettime(CLOCK_MONOTONIC, &t_start);
    rx_bytes = read_int_from_file("/sys/class/net/wan/statistics/rx_bytes");
    clock_gettime(CLOCK_MONOTONIC, &t_end);

read_int_from_file() basically consists of fopen(), fgets(), and sscanf().

Please note that I don't have kernel/driver programming experience so far.

Thank you for help and any suggestions!

Jan-Philip Gehrcke

--
http://gehrcke.de

#1 lore raw
Quoted 35 lines · Jan-Philip gehrcke

Dear list,

I'm about to write a small C program for measurements of the bandwidth on some network interface of my home router (uname -a: Linux fritz.fonwlan.box 2.6.19.2 #2 Thu Nov 18 16:35:17 CET 2010 mips GNU/Linux). I would like to accomplish very *precise* results, while keeping the absolute measurement time low.

In general, this problem requires the knowledge of time and byte differences. Therefore, two time measurements and two byte counter measurements for each, received and sent bytes, are required. Currently, I am measuring time via

clock_gettime(CLOCK_MONOTONIC, ...)

and reading the byte counters by reading files in sysfs, e.g.

      /sys/class/net/${interface}/statistics/rx_bytes.

This is very easy and works well, but reading these files four times takes some (varying) time* on my device, which introduces an error to my calculation.

Hence, I am wondering, if there is a way to receive this kind of data directly and much faster, via "kernel API". My question is related to this unanswered one from 2005:
http://linux.derkeiler.com/Mailing-Lists/RedHat/2005-02/0557.html

*On my device, reading such a file and interpreting the result as integer takes 1-10 ms, measured via:

     clock_gettime(CLOCK_MONOTONIC,&t_start);
     rx_bytes = read_int_from_file("/sys/class/net/wan/statistics/rx_bytes");
     clock_gettime(CLOCK_MONOTONIC,&t_end);
Quoted 8 lines

read_int_from_file() basically consists of fopen(), fgets(), and sscanf().

Please note that I don't have kernel/driver programming experience so far.

Thank you for help and any suggestions!

Jan-Philip Gehrcke

The netlink API can get stats for you, but I don't know if it would be any more accurate time-wise. It's not exactly trivial to implement this API, but you can look at the 'ip' program to get an idea:

git clone git://git.kernel.org/pub/scm/linux/kernel/git/shemminger/iproute2.git

There is also /proc/net/dev, but that may be worse than using sysfs as far as performance and time accuracy goes.

Thanks,
Ben

Signature · 2 lines
Ben Greear <[email protected]>
Candela Technologies Inc  http://www.candelatech.com