[LEDE-DEV] lantiq: reduced throughput with SMP disabled

[Felix Fietkau]

On 2016-08-04 08:11, James Finnie wrote:
> I've been chasing down some poor performance with my HH5A running LEDE
> (I've used r1142 - r1227 for testing).  Most notable when using
> multiple interfaces, particularly the combination of 2.4G wireless
> with VDSL at high throughput (am lucky enough to get 80 down / 20 up
> on my connection), resulting in very "bursty" performance during
> internet speed testing (thinkbroadband), and an inability to max out
> the connection (which was possible with stock firmware).  Brief
> inspection with top looked like there was no idle CPU available during
> such testing.  Some more in-depth tests using iperf3 [0] seemed to
> indicate that full rate VDSL and wifi was never going to be possible
> on this CPU.
> 
> Via a PM on said forum a suggestion came that it might be worth trying
> to revert the disabling of SMP (which was put in place due to an issue
> with dual eth interfaces).  I got a build environment up and running
> and locally reverted this change; it seems this solves my issues of
> CPU contention, pretty much saturating the VDSL over 2.4G wireless
> with CPU to spare, and results in quite a lot higher maximum
> throughput measured with iperf3 via both localhost (RX/TX 428/428Mbit
> vs 486/574Mbit) and ethernet (53/98Mbit vs 71/201Mbit).
> 
> I understand there are reasons to have SMP disabled; but it does seem
> that if like me you have fast connections then the penalty for this is
> pretty high due to the CPU choking.  It would be nice to get SMP
> re-instated if possible.
Please test the latest version of my staging tree at
https://git.lede-project.org/?p=lede/nbd/staging.git;a=summary

- Felix