On 14 Dec. 2016, at 22:42, Andreas Krey a.krey@gmx.de wrote:
On Wed, 14 Dec 2016 21:43:28 +0000, teor wrote: ...
The bwauth calculations do take latency into account, and they should: if CPU usage or bandwidth are near their limit, the latency through the relay will be high.
I stand corrected.
I observed my relays (a few years ago) to often run into the bandwidth limit, aka 'flatlining', and this having latency. I then started to set lower advertised bandwidth, and this went away. Problem here is that these are short-term event in relation to the bandwidth probes, so the probing can't really control this.
...
This has the drawback that relays located away from the US/Western Europe get poor scores.
What kind of latencies are we talking about here? And how much latency makes up for what bandwidth?
That's a really good question, the following factors interact: * bandwidth, * network peering, * latency.
I don't know the answer, but in my experience, an Exit relay located in France picked up bandwidth extremely quickly (peak in 4-5 months), one in Canada was a little slower (6 months and still at 60% or so) and some non-Exit relays in Australia are operating at 10% of capacity. (They don't all have the same bandwidth, so it's not a fair comparison.)
I'd have to ping each of the bwauths from each relay to be sure of the latencies, but as an example:
To moria1: France: 94 ms Australia: 245 ms
To gabelmoobwscan: France: 15 ms Australia: 340 ms
Note that it takes several round-trips to set up a TCP connection, and more round-trips to set up a Tor circuit (and acknowledge cells).
T