Remote Windows server for visualization

[Cardone, Giovanni]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hello everyone,

a couple of years ago we set up in our facility a computing node with Windows Server installed, which our users can access remotely and concurrently for image processing tasks, using either open source or commercial software. We have been very happy with it, and we are planning to add a graphic card to it, in order extend the functionalities of the node to visualization and GPU processing. Theoretically the officially supported cards are the NVIDIA M10/M60/P40/P100. We were wondering if anyone had direct experience with a set up like this, and if there are known limitations to better know in advance. We were thinking mostly of possible issues with Remote Desktop connection configuration, cards compatibility and software license management. Any feedback would be greatly appreciated!

Thanks,
Giovanni

[George McNamara]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hi Giovanni,

NVidia Titan RTX GPU has 24 Gb ram for $2500 ... at ~16 Teraflops
(single precision).  And will let you ray trace if/when you have
software for it. Can always be moved to some other workstation later if
you need the graphics resolution.

By coincidence (?) a new ranking was posted by Tom's Hardware today,
Titan RTX scored best

https://www.tomshardware.com/reviews/gpu-hierarchy,4388.html

GPU Performance Hierarchy 2019: Video Cards Ranked
by Chris Angelini March 25, 2019 at 12:00 PM

enjoy,

George

On 3/25/2019 11:12 AM, Cardone, Giovanni wrote:

> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Hello everyone,
>
> a couple of years ago we set up in our facility a computing node with Windows Server installed, which our users can access remotely and concurrently for image processing tasks, using either open source or commercial software. We have been very happy with it, and we are planning to add a graphic card to it, in order extend the functionalities of the node to visualization and GPU processing. Theoretically the officially supported cards are the NVIDIA M10/M60/P40/P100. We were wondering if anyone had direct experience with a set up like this, and if there are known limitations to better know in advance. We were thinking mostly of possible issues with Remote Desktop connection configuration, cards compatibility and software license management. Any feedback would be greatly appreciated!
>
> Thanks,
> Giovanni

[Olaf Selchow]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hello Giovanni,

From my experience,

- Visualization over a remote desktop requires proper configuration of drivers so that the rendering happens on the server GPU. By default, most drivers pull the data to the client / RDP terminal and render locally. This can lead to very poor performance due to limited network bandwidth and / or client GPU power. I've been working with some labs and companies in the past and it seems that the appropriate GPU setting work best on Quadro GPUs. Actually we haven't been successful to realize 3D renderings etc with anything else but Quadro GPUs. No success with Teslas or GeForce (doesn't mean it's impossible, but most likely won't work out of the box). Unfortunately, Quadro GPUs are expensive ... I know. The good news is that with such a configuration installed successfully, it's a great setting: You can use the full hardware power of the server from your server while sitting with your laptop in the cafeteria :-)

- GPU processing: Windows Server doesn't make your life easy when it comes to hardware resource management. a) few software packaged allow you to choose a GPU for your processing job (SVI Huygens does ... and also CARE DL "content aware image restoration" ... through a python -> tensorflow.) Other software just asks windows driver software for GPU ressources, and gets a handle for GPU #1 ... independent of how many GPUs you have installed. b) when multiple users need hardware ressources (GPU) the only reliable way I know, to give one user more or specific GPUs is to work with virtual machines. Windows on-board tools don't allow to distribute this efficiently. If you need more info on virtual machine environments, the microscopy facilities from Biozentrum Basel and Univ. of Zurich have a lot of experience with this!

- Licenses: Many commercial software providers will ask you to buy a floating license server (even with a single license). But it ususllay works great then - I have been working sucessfully with multiple concurrent users on a Windows Server using Imaris, arivis Vision4D, Fiji, Leica LAS X, Huygens and many more, even using 3D renderings and analysis. There are some limitations with Zeiss ZEN and Nikon NIS Elements: They can be launched only once per operating system. Means: User 1 gets the software, User 2 get the message "Software already in use". Main reason: USB dongle license. There is technical work-arounds and solutions from these companies, but they are not (yet?) sold officially.

I hope this helps.
Best of success!
Olaf


** Disclaimer: I work as an independent consultant, but in some projects I work with commercial providers of hardware and services in the field of Microscopy & BioImaging, Image Procesing & Large Data Solutions. This post is written by me with no commercial interest.

———
Olaf Selchow
Microscopy & BioImaging Consulting
[hidden email]