Re: Torperf

[Karsten Loesing <karsten.loesing@xxxxxxx>]

On 10/17/10 12:19 PM, Sebastian Hahn wrote:
> On Oct 15, 2010, at 4:13 AM, Mike Perry wrote:
>> Thus spake Karsten Loesing (karsten.loesing@xxxxxxx):
>>
>>> - Do we want to keep the #1919 Torperf runs running or migrate them
>>> to some other VM (that has enough memory)? What do we expect to
>>> learn from keeping them running or migrating them that we didn't
>>> learn from the first week or two? Instead of keeping them running
>>> we could also make a PDF report and put it on
>>> metrics.tpo/papers.html.
>>
>> I think this is very important to keep running, and that we should
>> think about adding new runs for based on the ratios of measured
>> consensus bandwidth to published descriptor bandwidth. Guards with
>> high ratios for this value have been observed by the bandwidth
>> authorities as having lots of slack capacity, where as Guards with
>> low ratios would be overloaded.
>>
>> I would love to have all of these datastream available for comparison
>> when various events and perf tweaks change the network. In fact, I
>> would love it if we could have the following 5 torperf runs logging
>> continuously and all overlayed on the main Torperf metrics graph:
>>
>> 1. Fastest 3 guards by network status
>> 2. Fastest 3 guards by ratio of ns_bw/desc_bw
>> 3. EntryGuards=0 (default current torperf)
>> 4. Slowest 3 guards by network status
>> 5. Slowest 3 guards by ratio of ns_bw/desc_bw

Okay, I think we can continue running the 3 or 5 Torperf runs with
modified guard node selection algorithms in the future if we think the
data is useful. However, note that

a) the 3 Torperf runs for #1919 run on some machine which is, AFAIK, not
one of our VMs and which may or may not host the 3 or even 5 Torperf
runs in the future,

b) I have no idea how to add 2 more Torperf runs with the guard
selection algorithms you stated, and

c) we're not using the output data for anything yet.

Mike and Sebastian, any ideas about a) and b)?

I think I can help with c) once I have the data.

>> Is it hard to keep all of these running and logging for some reason?
>> Does the 1919 script take up a lot of RAM?
> 
> The problem wasn't that the script was taking a lot of RAM, but rather
> that each torperf instance comes with its own tor instance. Spinning
> up another 9 clients caused the memory issues.

I think we can ask for more memory for the VM running this.

>> I can make these and any other changes to the 1919 script to help this
>> along.
> 
> So far it appears the script is doing fine.

Does that apply to the additional 2 Torperf runs, too?

Best,
--Karsten