[Author Prev][Author Next][Thread Prev][Thread Next][Author Index][Thread Index]

Re: Tor client performance



Roger Dingledine wrote:
> On Fri, Apr 23, 2010 at 08:21:15PM +0200, Olaf Selke wrote:
>> Roger Dingledine schrieb:
>>> Take a look at
>>> http://metrics.torproject.org/torperf-graphs.html
>> how can one explain the leap in latency two weeks ago?
> 
> That one's easy. We tried turning on the new "CircPriorityHalflifeMsec"
> feature, which gives priority to cells from circuits that haven't sent
> much lately. But there was a bug in parsing it from the consensus, which
> caused relays to pick totally the wrong cells, starve some circuits, etc.
> Bad news. So a week or so later we fixed the bug, and turned off the
> feature.
> 
> The question I can't answer is "how can one explain the continued horrible
> latency even after you turned circpriorityhalflifemsec off?"

is blutmagie exit used for latency measurement? I'm not sure about the
increase in latency being related to my futile attempts reducing the
number of tcp sessions in mid of April. But maybe recent decrease might
be related to blutmagie hardware upgrade at 5th May. The heavily loaded
dual core cpu has been replace by a now moderately loaded quad core.
Besides that it provides some more bandwidth with less load on each
core. Pls have a look at http://metrics.torproject.org/torperf-graphs.html

According torperf data from
http://metrics.torproject.org/csv/torperf.csv latency started to drop
5th May:

source		date		q1	md	q3
torperf-50kb	02.05.2010	7735	13092	20525
torperf-50kb	03.05.2010	9254	14550	20619
torperf-50kb	04.05.2010	7810	12498	20961
torperf-50kb	05.05.2010	6459	9848	15739
torperf-50kb	06.05.2010	4852	8069	14107
torperf-50kb	07.05.2010	4012	6036	10384
torperf-50kb	08.05.2010	3859	6080	9807
torperf-50kb	09.05.2010	3733	5766	9492
torperf-50kb	10.05.2010	3417	5073	6876

Was it me?

regards Olaf