Thanks for running the tests. Which exit nodes led to poor performance? I would like to try to reproduce any performance problems.
I did not record the nodes (they were in Europe). A simple test you could run on your server is fetching directory info from nodes that have directory functionality enabled.
wget http://<relay IP>:<dir port>/tor/server/all
e.g.: wget http://176.126.252.11:443/tor/server/all
You can get a bandwidth-sorted list of nodes at: https://torstatus.blutmagie.de/
There is a column that has the directory port.
How would you measure performance between my node and a given exit without being influenced by the properties of the middle relay? You can only set me as an entrynode, and you can't pick a specific middle,
You are wrong. :-)
You can build arbitrary circuits by hand. There are libraries like Stem, Txtorcon, TorCtl and there is a text based, SMTP-like protocol that you can use directly:
https://gitweb.torproject.org/torspec.git/tree/control-spec.txt
so how would you know that the low performance was my node and not the random middle relay?
Your node would be a non-random middle relay.
The bandwidth auths probably downrate the measurement results of your server severely because of those slow connections.
Probably? How can we investigate further?
AFAIK, the raw bandwidth auth measurements are not published, only the total result.