[metrics-team] TorPerf graphs

[Karsten Loesing]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256

On 18/12/16 17:56, David Goulet wrote:
> Hello everyone!

Hello David,

> This morning I noticed this change in the torperf graph on
> Metrics:
> 
> http://rougmnvswfsmd4dq.onion/torperf.html?start=2016-06-18&end=2016-12-18&source=all&filesize=50kb
>
> 
http://rougmnvswfsmd4dq.onion/torperf.html?start=2016-06-18&end=2016-12-18&source=all&filesize=1mb
> 
> I put there at 6 months interval to see the _huge_ change in
> download time. Same goes for the failure rate:
> 
> http://rougmnvswfsmd4dq.onion/torperf-failures.html?start=2016-06-18&end=2016-12-18&source=all&filesize=50kb
>
> 
http://rougmnvswfsmd4dq.onion/torperf-failures.html?start=2016-06-18&end=2016-12-18&source=all&filesize=1mb
> 
> This seems to have started in mid-October and the only thing I can
> see there is 0.2.8.9 being released with a security fix...
> 
> No surge of users or any change in bandwidth of the network as
> well:
> 
> http://rougmnvswfsmd4dq.onion/userstats-relay-country.html 
> http://rougmnvswfsmd4dq.onion/bandwidth.html
> 
> Did anything changed on Metrics side around that time? What Tor
> version is being used for the client of Torperf? This is worrying
> me as a potential issue that has been introduced recently
> destabilizing the network.

Okay, I started looking into this and found something.  Here's a graph
I just made with some hand-written annotations:

https://people.torproject.org/~karsten/volatile/torperf-2016-12-19.pdf

Some observations:

 1. This is what we'd expect: most values should be within a certain
range, and this is also where we'll likely find the median.

 2. What the heck?  This pattern for sure looks suspicious.  We'd
expect a plot like the 5mb one just with more dots (because there are
more measurements).  And note that this is a log scale!  This looks
really strange.

 3. The 1mb graph has a similar pattern, though not as visible.  We
can expect the 5mb graph to have one, too, but that one is practically
invisible on this graph.

So, the spikes that you saw are misleading, as there were no specific
events happening on those days.  It's just that the third quartile
jumped to one of those values in the graph I posted, and then back the
next day.

But those spikes led us to something else that looks really strange.
I just wonder what it could be.

> If nothing changed and this could be a legitimate issue, we should
> double check with Rob's OnionPerf data.

I cannot analyze this this week.  If somebody else wants to analyze
this, possibly by comparing it to Rob's OnionPerf data, please go
ahead!  Let me know what you need from me.  Maybe the Torperf results
on CollecTor are a start.  Let me know if you need more.

> Thanks! David

Good catch!

All the best,
Karsten

-----BEGIN PGP SIGNATURE-----
Comment: GPGTools - http://gpgtools.org

iQEcBAEBCAAGBQJYV7EgAAoJEC3ESO/4X7XBd7UH/0dLL75X/2qfybGQJQuncTVb
97+tJPq+wYoPxRPqHg+++q8b1T2QL49FBhdkNkv+xhCL3LXh0rGGUemPsibVGHMl
NmHxTC60omukvSB11W/+AMdYZpyQGqVXXpKwOOSxkqEudDPdmEjMVNc/0G12DCGl
67O5R8z7qD0sddzPY9CWN2+h4rgNR3IbRj9nqhV00OJgrh5j1OCR4e5ycxYJobDJ
f1HYkxAeaBBCJbEqsGcqVkGrHRWTXtTzT20fQls5US1cppYMIFlnVTtyVuZRM8U5
r3pDdQcQLFDG56sITIs6WhmzIwKkJsk46VPWTauRHKA+VxkZQWuQbX/MmjQALAA=
=myq1
-----END PGP SIGNATURE-----