[fedora-arm] disable turbo mode on panda

Jon jdisnard at gmail.com
Mon Dec 10 21:07:03 UTC 2012


UPDATE.

Looks like we found a regression.

We used to have /etc/sysctl.conf:
vm.min_free_kbytes = 12288


We used to set this in our build script, but it might not have transitioned
to the new image creation tools.

So I take everything back, leave the turbo_mode alone, and put back
the minimum free memory.

Also, systemD ate my sysctl.conf  :(

Sorry for the noise.

-Jon Disnard


On Mon, Dec 10, 2012 at 1:30 PM, Jon <jdisnard at gmail.com> wrote:

> Hello team :)
>
> Panda-es, f18 beta
>
> Sometimes the board spews out lots and lots of these kevent 2 lines:
>
> [root at panda-f18-v7hl ~]# dmesg | tail
> [ 4325.019653] smsc95xx 1-1.1:1.0: eth0: kevent 2 may have been dropped
> [ 4325.026367] smsc95xx 1-1.1:1.0: eth0: kevent 2 may have been dropped
> [ 4325.033081] smsc95xx 1-1.1:1.0: eth0: kevent 2 may have been dropped
> [ 4325.039764] smsc95xx 1-1.1:1.0: eth0: kevent 2 may have been dropped
> ...
>
>
> Problem statement:
> Those kernel logs look unhappy.
>
> Impact statement:
> AFAIK this might result in some dropped packets.
>
> Steps to reproduce:
> Every time, just use wget to fetch a large file.
> I grab an f17 iso from my local file server.
> The console will start flooding the kevent, and the wget gets very unhappy.
>
>
>
>
> Supposedly disabling turbo mode on the cmdline (uEnv.txt) would help:
>
> smsc95xx.turbo_mode=N
>
>
> I suppose turbo mode is some kind of interrupt throttling scheme?
> It might reduce throughput while increasing resiliency?
>
>
> Tested with turbo disabled, and I did not see a significant reduction in
> throughput, or increased latency, but a dramatic increase in resiliency.
> aka disabling turbo_mod stopped the kevent 2 problem.
>
> I propose we make this default for panda and beagle.
>
> Your thoughts?
>
>
> --
>
> -Jon Disnard
> irc: masta
> fas: parasense
>



-- 

-Jon
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.fedoraproject.org/pipermail/arm/attachments/20121210/d85cd312/attachment.html>


More information about the arm mailing list