perfSONAR User Q&A and Other Discussion

["Kaufman, Ian" <>]

Have you tried disabling NetworkManager?

I have done so several times so that I could hand code interfaces for various reasons.

Ian

Ian Kaufman
Systems Integration Engineer

UC San Diego, Research IT Services

ikaufman AT ucsd DOT edu

---

From: <> on behalf of Mark Feit <>
Sent: Friday, January 12, 2024 8:59 AM
To: Masshardt, Eric <>; <>
Subject: Re: [perfsonar-user] Questions around Network Manager, EL 8/9

 

Masshardt, Eric writes:

 

Is system-config-network still present in EL 9? Sounds like I need to run through a “Server with no GUI” install of Rocky 9 somewhere to check things out.

 

It’s a hybrid for the time being.  The old-style files are still supported, but I’d bet dollars to doughnuts that EL10 will discontinue it.

 

We’re in the process of dropping support for distributions that reach EOL shortly and will be removing some of the baggage they’ve been carrying with them.  This will be a good opportunity to assess what changes can be made to bring networking into line with current practice.

 

One of our issues is movement of large amounts of data through our network. We wanted to see if PerfSONAR can assist in (1) on-going baseline of network performance between a number of key/different locations (2) help in troubleshooting and identifying the sources when we are experiencing issues.

 

Both of those are perfSONAR’s raison d’être.  There are a number of built-in features that will let you give tasking to a whole fleet of them at once (pSConfig), archiving measurements to a central location (the Archive bundle) and graphing the results.  If you have your own systems for these things, pScheduler, which oversees the measurements, can be configured to send results outside of perfSONAR using a number of different methods.

 

We have combinations of workstations and servers that consolidate data to be transferred and, if possible/reasonable, would like to utilize these for PerfSONAR as well. If we need to halt transfers to be able to effectively test, we could likely do that. These nodes currently have 10GbE (ports of which still seem to be a precious commodity) while most systems do not.

 

The recommended practice for production use is to put perfSONAR on a dedicated machine so the measurements aren’t distorted by other system activity.  Sharing for a proof of concept is fine if you can live with the side effects or mitigate them by stopping the other activities.  Sounds like you can do the latter, so you’re good there.

 

I run the perfSONAR installation on SCinet, where most of the test points are 10 Gb/s and are still good at identifying problem links.  In a 10 Gb/s environment, you may find that 1 Gb/s will do the same.  For that, you can re-purpose old desktops or servers as perfSONAR nodes and get useful service out of it.  In my organization’s network, we’ve started putting less emphasis on throughput and more on latency.  But we run a backbone where that works out better than in, say, a campus environment.

 

Sounds like hardware requirements aren’t overbearing for dedicated test nodes. Initial testing on the systems we have, even if non-optimal, will hopefully give insight into the utility and usefulness of PerfSONAR and then we could go from there. With that in mind, KVM based virtual would be better than Docker based install of PerfSONAR?

 

Given what you said above, either would work.  VM environments tend to be I/O-constrained, but I think KVM’s virtio solves that.  Running in a VM would also let you install the entire toolkit; the Docker container is just the testpoint.    

 

Hope that helps.

 

--Mark