Re: Tool for sampling /proc/net/softnet_stat statistics
From: Jesper Dangaard Brouer
Date: Tue Mar 08 2016 - 04:51:16 EST
On Mon, 7 Mar 2016 12:54:13 -0500
Willem de Bruijn <willemdebruijn.kernel@xxxxxxxxx> wrote:
> On Mon, Mar 7, 2016 at 10:36 AM, Jesper Dangaard Brouer
> <jbrouer@xxxxxxxxxx> wrote:
> > Hi Google,
> > While playing with RPS, I needed to read stats from
> > /proc/net/softnet_stat and the tools I could find  and  was not
> > very good.
> > I lack of better, I coded up my own tool softnet_stat.pl here:
> > https://github.com/netoptimizer/network-testing/blob/master/bin/softnet_stat.pl
> > The output format/columns in /proc/net/softnet_stat is undocumented,
> > plus values are printed in hex. E.g. to decode the columns you need to
> > read kernel function kernel softnet_seq_show() in
> > kernel/net/core/net-procfs.c.
> > To make things easier I wrote this small perl script for get
> > so human readable statistics from /proc/net/softnet_stat.
> Very nice. Thanks for sharing, Jesper. I maintained something similar,
> but never got around to clean up and upstream it. Will start using
> yours, instead.
> A few points, from using my earlier tool:
> A minimum cut-off value is helpful, especially on beefy servers, to
> suppress the many 0 rows. Preferably configurable, to also be able to
> suppress low-rate background traffic when analyzing a few large
> streams. My default was 500.
> The number of columns has grown with kernel versions. The latest
> column is flow_limit, added in 3.11 at99bbc7074190. It is helpful for
> the script to be robust against both older and future kernels. On
> which note, to be able to support these kinds of tools, any new
> columns to such procfs files should be appended, not inserted as for
> instance in https://patchwork.ozlabs.org/patch/574171/
> Time squeeze is an exception, in that it is number of squeeze events
> per second, not number of packets squeezed. This is often
> misunderstood if not explained clearly.
Thanks a lot for your feedback. I don't have time to address it right
away, so I've instead added a section with future development todo's.
So, I don't forget this valuable feedback :-)
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
Author of http://www.iptv-analyzer.org