By Tobias Kreidl
Note: This is part one of a guest series from Tobias Kreidl, an NVIDIA GRID Community Advisor (NGCA) and Academic Computing Team Lead at Northern Arizona University. Tobias has considerable experience in using a range of NVIDIA technologies for HPC, VDI and graphics as a long-time user of NVIDIA Quadro, Tesla and GRID products.
My First Personal View of NVIDIA GRID
I still recall experiencing the announcement of the direct integration of a virtual graphics processing unit (vGPU) into XenDesktop using the new NVIDIA GRID technology at Citrix Synergy 2013 in Anaheim. The demos were awe inspiring and the NVIDIA booth drew many gawkers who were treated to viewing real-time rendering of complex video with amazing resolution and speed. It brought back memories of the then very pricey Trancept TAAC-1 graphics accelerator board we had added onto a Sun Microsystems 4/260 workstation clocking in at 16.67 MHz back in the late 1980s. Being able to spin around 3-D models in real time and in color was phenomenal back then.
A fundamental difference is that now, GRID technology can be shared concurrently by multiple users, all tapping into a server from different locations, and not stuck waiting for a turn to sit down at a fixed unit with a single CRT monitor that weighs 100 pounds! It was rapidly clear to me that the NVIDIA GRID technologies were going to be something big and not restricted to those with deep pockets or constrained by limited applications that could leverage this technology. It was evident that this was just a beginning.
Fast-Forward to Mid-2016
Things have evolved and in particular with the latest NVIDIA GRID products, the hardware and software options have grown substantially. The latest Tesla GPUs (M6 , M10 and M60) offer much greater computational and memory capacities compared to the original GRID K1 and K2, including blade server support in the case of the M6.
Along the way, vendors have been able to tap into this technology, notably Citrix, VMware and Microsoft. The ability to configure both passthrough and dedicated GPUs makes for a great amount of flexibility and allows for right-sizing the allocation of resources to meet the needs and expectations of end users. Students, in particular on the undergraduate level, are not going to have the same demanding requirements as professionals in the CAD-CAM area. This makes possible higher user densities for student-oriented installations.
Another area of advancement has been that of video delivery. Without this many operations would not be possible with reasonable performance due to a combination of factors, such as limited bandwidth, packet loss, latency, and rendering constraints. Fortunately, many advances in this area have taken place and continue to do so including, Citrix HDX, VMware Horizon Blast Extreme and the ability to selectively leverage H.264 decoding.
In short, the increasingly more powerful clients have reached the point where even a Raspberry Pi 2 or 3 has sufficient capabilities to serve as a thin client and be able to handle the throughput generated by GPUs on the server back end. One such example can be seen in this video I made with an RPI 2 and here, a RPI 3 remote session to a cloud service with an NVIDIA GPU back end, which shows that even very modest clients are fully capable of dealing with high-resolution graphics.
Adopting the Technologies
NVIDIA’s own impressive virtualization case study archive shows how industry and manufacturing quickly embraced this technology. The ability to centralize data and work on complex graphics editing and display around the globe revolutionized the way projects could be undertaken even on the international level. Being able to allocate GPU resources in different amounts was also a huge benefit.
Without going into the myriad areas where GPU technology can be leveraged, an early realization was that there would be a need in the workforce for employees with the knowledge of how to leverage GPU technology. This is where educational institutions come into the picture.
Speaking from the perspective of my own institution, Northern Arizona University (NAU, https://nau.edu), we have seen the power that GPU acceleration can provide in particular with thin clients (a story we shared with Citrix, which you can read here) and specifically as of late in conjunction with the Citrix RPI HDX project. We have experimented with early GRID K1 and K2 GPUs over the last two years and are now launching a pilot that is based on the Tesla M60.
The initial research with GRID 1.0 was very informative as it allowed working with Citrix to better understand the desktop delivery mechanisms and codices used in conjunction with HDX. We had been ramping up our XenDesktop presence that was entirely based on thin clients running XenDesktop (now numbering close to 500 units). We had already been using thin clients for over a decade, being conscious of power-saving and other environmental impacts, as well as the advantages available with central management.
Three important things became evident during this process.
- First, only high-end thin clients could handle high-speed, high-resolution graphics, and it would be a losing argument to push for such expensive units versus conventional PCs. Another way had to be found to handle the video demands if thin clients were to remain viable options.
- Second, the ability to leverage GPUs on the server side not only provides more flexibility at the end-user side, but also opens up options for mobile user access.
- Third, the replacement cycles could be tailored very differently for devices that were cheap and easy to replace compared to much more expensive, static PC units. Thin clients can be used longer than conventional PCs and if they break or need to be repaired or replaced, the cost can be considerably less. Experiments with low-end thin clients with conventional as well as GPU-driven back ends were shown to be successful by carefully identifying which units were up for the task, avoiding the need to overspend unnecessarily. In particular for many of our basic lab and classroom needs, a Raspberry Pi with a single monitor is completely sufficient and significantly less expensive. In fact, replacing an RPI 3B board is less expensive than the cost of three years of firmware updates from a leading purveyor of thin clients. The cost savings can either be put into the server and storage back ends or into VDI licensing. There are also significant savings in the power needed to operate thin clients.
Working with GRID
Our current pilot will give us the flexibility to determine if GPU passthrough and/or dedicated vGPUs will best serve our purposes. Two Dell R730 servers, each equipped with dual Intel Xeon E5-2680 v4 2.4 GHz CPUs with 14 cores and 256GB of DDR4-2400MHz RAM and initially one M60 apiece, have been obtained. The configurations could readily be enlarged by adding an additional GPU and/or memory. We will look into multiple options. These include selectively accelerating applications on XenApp and terminal servers, providing specialized XenDesktop configurations to labs running thin clients, as well as remote access via BYOD.
The primary applications we are interested in include, but are not limited to, the AutoDesk suite, ESRI ArcGIS, Google Earth, and Adobe Creative Suite (Creative Cloud) and PhotoShop. We may also ultimately investigate cloud-based desktop and application options that are backed by NVIDIA GPUs. We may also incorporate other NVIDIA GPUs, such as various models available in the NVIDIA Quadro professional workstation line, as passthrough options for some of our physical terminal servers. The Kepler and Maxwell series offer excellent performance where non-virtualized GPU needs are required.
There is a huge amount of interest in leveraging GPUs in education. The potential for students, faculty and staff to tap into such resources is enormous. My goal for some time has been to advocate for the educational community in all aspects of IT. My roles as a member of the Citrix User Group (CUGC, https://mycugc.org/) Steering Committee, the Citrix Technology Professional (CTP) group, and even more recently as a member of the NVIDIA GRID Community Advisor (NGCA) program help serve this purpose. NVIDIA was responsive about revising its GRID pricing and has a dedicated program that provides significant discounts for GRID products to educational institutions.
The key to success in educational institutions, which are nearly all limited in funding, is to work with corporations who can help them remain on the cutting edge of technology through donations and discount incentives. As mentioned more than once earlier in this article, the benefit of giving students hands-on experience helps to give them the real-world skills demanded and sought in the workplace and hence, the benefits are mutual. Investment in education, as much of a cliché as it may be, brings very real, tangible results. We have seen our student workers find gainful employment in many excellent areas due to their experiences working with us on real-world projects, so they are both technically and mentally prepared to tackle challenging tasks.
I plan to continue this advocacy that extends to other companies and appreciate NVIDIA’s interest in the educational sector. We have had many good discussions on these topics. Aside from numerous private correspondences, I am active on numerous forums and social media outlets, have posted articles on several blog sites, was on the Birds of a Feather educational panel at Synergy 2016 as well as served as host at one of its educational table tech chat sessions. I believe in the future of these technologies and will continue to strive to help make them available to as many as possible.
In part two of my post, I’ll highlight more use-cases from universities around the world.
TOBIAS KREIDL, Ph.D.
ACADEMIC COMPUTING TEAM LEAD, NORTHERN ARIZONA UNIVERSITY
Tobias spent 15 years as a research astronomer before a career charge that took him to NAU’s (Northern Arizona University) central IT department. A strong advocate for the use of computers in education, he has helped introduce a number of technologies to students and staff, embracing many open source tools, as well as Citrix and other commercial products. An early advocate of virtual computing, he is a leading poster on the Citrix discussion forum and works with his team on everything from virtual servers and thin clients to software defined storage, GPU technologies, disaster recovery, Web-based services, and anything else that can enhance and secure the educational computing environment. Tobias brings considerable experience in using a range of NVIDIA technologies for HPC, VDI, and graphics as a long time user of Quadro, Tesla and GRID products.
To learn more about the NGCA and how you can become an advisor, click here.