[Babel-users] sending rtt always on 1.8?

[Dave Taht]

On Mon, Mar 28, 2016 at 12:03 PM, Juliusz Chroboczek
<jch at pps.univ-paris-diderot.fr> wrote:
>> what harm would there be to always timestamping, rather than
>> optionally, in the 1.8 release?
>
> Arguments in favour:
>
>   - Dave likes it;
>   - it wouldn't harm to add it to the default metric.

Heh. Eyeballing the stat via babelweb I can pretty easily determine
the media type (ethernet, wifi, wifi with powersave) and the workload.
Yes I retain hope for short rtt metrics, one day, after we get wifi
more fixed.

>
> Arguments against:
>
>   - it requires two extra calls to clock_gettime for each packet sent,
>     which is expensive on MIPS before kernel 4.5 and musl 1.1.13;

I didn't know that. in glibc at least it's a read from a universally
shared page - taking ns - rather than a syscall.

Is it really a 4.5 issue? (got link?). I thought openwrt was shipping 4.4, also.

Is it possibly an arm issue (most of what I'm working on is arm)?

>   - it adds 8 octets to each HELLO and IHU TLV;

mmm... could use a spreadsheet to calculate other overheads... got a
situation where every box I just played with sends 6 p2p routes by
default...

>   - it might cause oscillations in topologies we didn't envision;

I was not suggesting turning it on as a metric at this point, merely
having it on.

>   - it adds another moving piece;
>   - it's not an RFC.
>
> So I'm hesitating.  I'll probably wait until 4.5 and 1.1.13 are in OpenWRT
> and have been deployed.

I can wait.

> -- Juliusz