perfSONAR User Q&A and Other Discussion

["Chevalier, Scott S" <>]

I think Mark and Marcos made some excellent replies already and shared some interesting examples, but if anyone is looking for longer form explanation of why caution should be used when relying on VM architecture I’ve provided links below to the perfSONAR Docs reference page.

pS Docs reference: https://docs.perfsonar.net/install_virtual_machine_details.html?highlight=virtual%20machines
ESnet Faster Data VM tuning: https://fasterdata.es.net/host-tuning/linux/virtual-machines/

 

Cheers,

 

Scott Chevalier

Network Systems Analyst

International Networks @ IU

812-856-9964

 

From: <> On Behalf Of Marcos Felipe Schwarz
Sent: Wednesday, July 1, 2020 2:31 PM
To: Mark Feit <>; Paul Manomaitis - NOAA Affiliate <>;
Subject: [External] [perfsonar-user] RES: Greetings, questions on PS Software and Hardware

 

This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources.

 

Just complementing on virtualization drawbacks. We are currently working with 1G/10G perfSONAR measurement nodes on ESXi. But the performance with default configurations is really poor. Specially drift between VM and host clocks which affects one way latency tests, and overall the system is less performant due to CPU schedulling, dynamic RAM allocation and network virtualization.

We tuned the hypervisor and the VM and got results similar than baremetal, removing scheduling of resources, dedicating/preallocating CPU and RAM and using NIC passthrough.

 

Follows the documents we used back than, maybe they are a little bit outdated.

https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/techpaper/latency-sensitive-perf-vsphere55-white-paper.pdf

http://download3.vmware.com/software/vmw-tools/papers/VMTJ_issue_4.pdf

 

Regards,

Marcos Schwarz 

 

De: <> Em nome de Mark Feit
Enviada em: terça-feira, 30 de junho de 2020 18:01
Para: Paul Manomaitis - NOAA Affiliate <>;
Assunto: Re: [perfsonar-user] Greetings, questions on PS Software and Hardware

 

Paul Manomaitis - NOAA Affiliate writes:

 

1. PerfSONAR is best used in the downloadable CentOS download package, this will have the greatest level of testing and assumed stability? vs 

 

Whether you install from the ISO or start with a minimal CentOS, you’re largely running the same thing.  If you install the toolkit, that does a lot of the basic tweaking for you.

 

2. Are there any drawbacks to virtualization? Using VirtualBox or VMWorkstation running on a windows OS.  Or even running it as an ESXI guest?

 

The big drawbacks are handing control of the system over to something else and I/O throughput, which tends not to be wonderful on VMs.  Some of that can be mitigated by dedicating resources and doing things like PCI passthrough for the NICs.  Really, though, the only way to see whether a given setup will get the job done is to run it in the lab back-to-back with a machine whose performance you know.  All that said, there are a lot of registered perfSONAR instances running on VMs, and while I doubt they’re doing high-speed throughput, they’re just fine for most other things.

 

3. For bare metal, is there an optimized rack mount / non-rack mount minimal footprint compute solution? e.g. An intel NUC, since Raspi/Odroid with ARM isn't supported. -- A HW loadout which might be recommended ? Emphasizing 10G NIC capability.

Trying to balance remote management suite capabilities and size of footprint and cost.

 

If you’re after a box to do 10 Gb/s, Ed Colone at UMich is a big fan of the SuperMicro E300-8D.  It’s rackmountable, well-equipped and relatively inexpensive (not NUC/RPi cheap, but a lot of machine for the money).  He has a few notes on it here:  https://github.com/perfsonar/project/wiki/Small-form-factor-10GE-perfSONAR-testpoint.

 

--Mark

 

Attachment: smime.p7s
Description: S/MIME cryptographic signature