On 12/07/2012 16:36, Ian Collier wrote:
> OK. I will try this one more time then I will shut up.
>
> The latency tests will not adversely affect the bandwidth tests - so at that level they will at least not affect the agreed purpose of the boxes.
>
> However, at any time bandwidth tests are running a: the network interface will be to some degree be saturated and b: the cpu  may also be busy ( b: is especially true for 10 gig ). Both of these will tend to firstly increase observed latency and secondly cause some amount of packet loss.
>
> This is nothing to do with perfsonar per se, just about running iperf and similar tests - or indeed just using your network a lot. (You'd see the same effects on busy disk servers.)
>
> As you increase the number of hosts you test against, you also increase the amount of time your bandwidth tests are running. So you will also increase the amount of time any latency tests are being distorted. (So at the moment, testing against just a handful hosts you might not see a large effect.)
>
> What all this means is that your latency tests will be telling as much if not more about how busy your bandwidth test hosts  are rather than the quality of the network between your hosts.
>
> Internet2 and the perfsonar developers know what they are talking about when it comes to network monitoring. They make good recommendations. I don't understand this enthusiasm for doing what they explicitly suggest you do not do.

My tuppence-worth from a position of blissful ignorance.

 From http://psps.perfsonar.net/toolkit/FAQs.html#Q46
"
Q:Can I run owamp and bwctl on the same host?
A:

bwctl testing will affect owamp results, so it is best to run them on 
separate hosts. If this is not possible, owamp is still a very useful 
troubleshooting tool, and it is better to run them both on the same host 
than to not run owamp at all.

If you do run owamp and bwctl on the same host, owamp may see increased 
latency and jitter during the bwctl tests, and on a 1G host you may even see 
loss. If you are using nagios to generate alerts based on these owamp 
results, you'll need to take this into account.
"

This would lead me to believe that the perfsonar developers themselves 
recommend running both services on one machine *if* you don't have two to 
run them separately.

Obviously the bandwidth tests will affect the latency/loss results, but if 
these tests are regular then I would expect this effect to be a background 
against which real loss or latency changes would be visible. Without running 
a PS box I couldn't say if that was true or not but I would consider it 
worth investigating for those sites that only have one box. (We have two so 
I have no axe to grind here).

John

> --Ian
>
>
> On 12 Jul 2012, at 14:24, Peter Gronbech wrote:
>
>> I agree with Ewan, in that I thought that was the plan agreed at the ops meetings.
>> I'm not an expert on perfsonnar so do not know if running both tests would be a problem, but it was not the intension.
>>
>> Cheers Pete
>>
>>
>>
>> --
>> ----------------------------------------------------------------------
>> Peter Gronbech  GridPP Project Manager          Tel No. : 01865 273389
>>                                                 Fax No. : 01865 273418
>> Department of Particle Physics,
>> University of Oxford,
>> Keble Road, Oxford  OX1 3RH, UK  E-mail : [log in to unmask]
>> ----------------------------------------------------------------------
>>
>>
>>> -----Original Message-----
>>> From: Testbed Support for GridPP member institutes [mailto:TB-
>>> [log in to unmask]] On Behalf Of Ewan MacMahon
>>> Sent: 12 July 2012 10:44
>>> To: [log in to unmask]
>>> Subject: Re: perfSonar machine in Sheffield
>>>
>>>> -----Original Message-----
>>>> From: Testbed Support for GridPP member institutes [mailto:TB-
>>>> [log in to unmask]] On Behalf Of Ian Collier
>>>> Sent: 12 July 2012 10:37
>>>>
>>>> You don't want bandwidth and latency tests running on/against the same
>>>> machine. (The bandwidth tests running are likely to make the latency
>>>> results meaningless.)
>>>>
>>>
>>> Indeed, I thought the plan was for the original six sites with
>>> a pair of boxes to run both, and the sites with just the one
>>> new machine (that was going to be a gridmon node) would now be
>>> set up as a PerfSonar bandwidth box, and those sites just wouldn't
>>> have latency measurements.
>>>
>>> In other words, I think Elena's current configuration is what I'd
>>> expect it to be.
>>>
>>> Ewan
>


-- 
John Bland                       [log in to unmask]
System Administrator             office: 220
High Energy Physics Division     tel (int): 42911
Oliver Lodge Laboratory          tel (ext): +44 (0)151 794 2911
University of Liverpool          http://www.liv.ac.uk/physics/hep/
"I canna change the laws of physics, Captain!"