Can a Virtual Workstation truly provide an experience comparable to a traditional deskside machine?

Posted by Alex Herrera on February 7th 2017 | Discuss
Categories: Hardware Review,
Tags:

Giving Boxx’s Pro VDI a test drive

Business executives and IT administrators alike have long been enamored with platform virtualization technology, and rightfully so. the advantages of hosting virtual desktops remotely in the data center are well-known and precisely the reason VDI — Virtual Deskstop Infrastructure (VDI), one but not the only example of hosted virtual desktops — has grown from nothing to a near $6 billion dollar business.

With the relatively recent advent of GPU-accelerated server-hosted virtual machines, businesses that rely on traditional workstations (deskside towers and mobile laptops) can finally join the virtual hosting opportunity. Many are in the process now of evaluating virtual workstation solutions to determine if they might replace or complement a business’s existing traditional workstation infrastructure.

But users don’t have the same interests and computing environments that executives and administrators do. Whereas the latter are concerned with security, cost, management and other business issues, the former’s concerns are only about doing their jobs effectively and efficently. And that means running their critical applications and workflow as effectively on a remotely hosted virtual machine as they’ve done it with traditional physical machines at their desk. If they can do that, great. But if they can’t, then all those promised benefits of virtualization are moot.

And that’s why we were intrigued to test a new product from Boxx Technologies, a server optimized to host virtual workstations to serve professional-caliber, graphics-intensive applications. Could it deliver a workstation experience comparable to traditional deskside machines? 

Boxx aggressively exploring virtual workstation technology with multiple hardware options

A respected long-time vendor of high-performance workstations for CAD and Media/Entertainment applications, Boxx  has flourished in the face of Tier 1 workstation competition (i.e. HP, Dell and Lenovo) by pushing the envelope on technologies  and new computing approaches. On the latter front, the company is leaving no stone unturned in its attempt to be a  leader in the emerging arena of virtual workstations.  

The company has introduced not one but two Boxx-branded server lines equipped with GPU accelerators and optimized  for virtual workstation hosting: Pro VDI and GRID. The GRID brand product delivers an official Nvidia GRID vGPU solution  running on Tesla GPU-based boards, while Pro VDI products a built on Quadro professional graphics boards. Why  both a GRID compliant and non-GRID product line? Well, the answer is two-fold. Some feel that dedicating a physical GPU  to a virtual machine (VM; i.e. vDGA) performs better than a VM with GRID vGPU running a dedicated (single-user) profile.  

The second answer is likely the more important one: Nvidia GRID licensing. Running an official GRID solution on top  of Tesla M-series GPUs (and future GPUs, presumably) requires one of three GRID license options. The appropriate license  for professional applications is the Virtual Workstation license, and that’s substantially more expensive — expensive enough  for many to pass on GRID and stick with the existing vDGA solutions running on Quadro or AMD Radeon Pro boards.  

Boxx Pro VDI series, including the only liquid-cooled, virtualizable server

Boxx’s Pro VDI series targets power users running applications like Solidworks and Maya, proposing a virtual machine capable of delivering an experience that engineers and designers would expect from their traditional deskside machines. represents the true VDI version of the XDI product line, comprised of rackmount servers incorporating multiple graphics boards (Quadro today, but has included FirePro / Radeon Pro). As is becoming the norm, products are being shaped (wisely we believe) with recommendations on minimum resources (per VM) and specific caps on user consolidation. Boxx’s current flagship Pro VDI product is the 3U Pro VDI 8401R-V.

One thing that separates Boxx’s Pro VDI GPU-enabled server products from its rivals is the same as what separates its deskside workstations from those of the competition: liquid-cooling. The company managed to overclock its 10-core Intel Core i7- 6590X (Extreme Edition) CPU from the nominal 3.0 GHz all the way up to 4.2 GHz. 

The virtual workstation experience: a Boxx Pro VDI test drive

Testing a virtual workstation on a Pro VDI server gave us the opportunity to measure a well-provisioned virtual workstation compared to a conventional workstation, while getting a feel for its interactive response as well. But to get to that point of running applications and benchmarks, we first had to get our hosted VM up and running—and that process happened surprisingly fast and error-free.

Boxx provided us with access to a GPU-enabled VM on a Pro VDI 8410R-V at the company’s HQ in Austin. We dow loaded and installed the VMware Horizon client to run on our local machine, and after a few steps to setup the host and entering user/password we had a familiar Windows 10 Professional desktop within minutes. It looked identical to a local, client-side desktop, with only a couple exceptions, such as special hooks for I/O (e.g. the client’s physical USB input has to be exposed to the remote server) and things such as Ctrl-Alt-Delete (which has to be differentiated as a command for server or client). Over a few days, we brought it up and took it down, quickly and solidly every time, and we think worth noting — it had no trouble discerning or properly displaying on a triple-monitor setup. 

Judging performance: throughput vs. interactivity

With physical specifications comparable to a well-equipped deskside tower, we’d expect the Pro VDI 8410R-V host to deliver visual performance comparable to a traditional deskside tower. Hence, running a benchmark like Viewperf or SPECwpc wasn’t likely to be particularly revealing.

Still, we were curious as to whether our workstation-VM running on that host, combined with any additional remoting overhead— in particular video stream encoding or deficiencies due to shared server resources might result in material differences in the graphics throughput of the local/physical model and the remote/virtual one. It’s worth nothing that while the Pro VDI 8410R-V host can support a hardware encoder for PCoIP protocol (the standard Teradici protocol employed by VMware to transport the video stream from server to client), our specific host did not have hardware support. As such, we’d expect some, albeit minor, overhead associated with PCoIP encode.

To confirm comparable (or not) performance, we ran Viewperf 12 on our Pro VDI hosted virtual workstation and compared results with a high-performance, traditional Boxx workstation with the same, dedicated GPU: the Quadro M2000. The comparison desktop workstation is also a Boxx machine, a recent (but not quite current) Apexx 4 machine with specifications quite similar to those provisioned for our Pro VDI VM.

For the most part, the results confirmed our expectations; the two machines delivered roughly the same scores. Interestingly, however, the higher scores came from the virtual hosted workstation. It’s not a huge difference, and we’ll chalk up the ~10 – 15% edge to the host’s slightly faster CPU, as well as more recent Quadro driver running on the newer OS. Furthermore, the fact that the VM outpaced the traditional deskside also leads us to believe that the PCoIP overhead is minimal. (For the record, that is what we’d expect, as PCoIP hardware encode is typically used for high-density user environments).

While 3D graphics throughput was more than respectable, it was not our primary area of concern with this evaluation — response time was. Remember, a remotely hosted virtual workstation has to rely on one critical component a deskside machine doesn’t (at least not during interactive periods): the internet. Computation and rendering happens on the other side of a network — perhaps a LAN but also a WAN (especially for cloud and corporate satellite offices) — and the user’s experience depends on that network’s capabilities, namely bandwidth, latency and the ability to adapt to congestion and delays across that network. In particular, it’s the round-trip latency that matters, the delay from when the user rotates a model view with the mouse, for example, until that new view is updated on screen.

So what we were far more interested in judging was whether or not the round-trip latency of going back and forth to the server across the WAN would result in slower, sluggish response. Now, “sluggish” is of course subjective, but many will quote 250 ms as a worst-case threshold, including Citrix. Exceed that number, and Citrix will actually recommend against using a remotely hosted VM, as the experience gets frustrating and work downright unproductive. And even 250 ms is not close to ideal, as delays become quite noticeable at that level.

To evaluate responsiveness, we relied not upon benchmarks but by rendering a Solidworks model on our virtual workstation and using the mouse to rapidly rotate, pan and zoom. We employed a fairly simple 50 MB model of a car, but again, the render performance was not the point. Rather, we were looking to see — rather, more to “feel” — how responsive the virtual workstation’s graphics can respond to our mouse movements. How did that turn out? Quite well, actually. While we could certainly notice that the response was not quite as snappy as a traditional local workstation, it was far from annoying or problematic. We had no way to measure, but based on past experience, our latency was probably in the ballpark of 100 ms or so, far from that 250 ms threshold. 

What do we think?

Few will debate the financial and logistical advantages of any remote, datacenterbased virtual hosting environment. The ability to better handle anywhere/anytime access for an increasingly global workforce, while more effectively managing big data and security with a remote server-approach paints a very compelling picture. However, whether those advantages can be realized by the professional class of workstation users depends on one simple question: can it deliver the experience designers, engineers, and others are accustomed to with their familiar deskside and mobile workstations? If it can, then it is a real benefit, and traditional workstation users everywhere will at least want to consider a move to the virtual workstation, either in whole or in part. But if not, all those merits everyone lists are pointless.

Granted, the answer to that is subjective, but by getting a chance to use such systems in a real-world environment give us all a chance to come to our own conclusions. We’ve checked out Nvidia’s GRID test drive, and even better, we’ve now had a chance to check out Boxx’s full-fledged virtual hosting environment, built on VMware VDI and running with a dedicated Quadro M2000 GPU. If any experience should give a reasonable indication, this is it.

We came away thinking the answer is yes, a virtual workstation can be created and provisioned to deliver the graphical computing throughput workstation users demand. More importantly, the round-trip client-server latency is reasonable, if not a “local” response, it was still well within what we think the majority of users would tolerate—however, it will vary depending on the network (e.g. number of hops) between server and client. Finally, the advantages of centralized, remotely hosted client computing can be realized not just for email and productivity applications, but hard-core workstation. The virtual workstation is real. But going virtual has its considerations, by far the biggest being network capabilities. If the network in use can’t support a reliable range of bandwidth and a reasonable worst-case latency, then forget the idea of virtual workstations. On the positive side, it’s worth pointing out that we took no special care in establishing any kind of special network access. Rather, we ran our client up in the California mountains, through a local cable provider to who knows where and eventually to Boxx’s Austin campus. And response time was acceptable. Put together a comprehensive, high-performance, purpose-built network for your business — and we’ve seen some impressive examples — and there should be nothing standing in the way of at least considering the prospect of ditching the physical and going virtual. – A.H. 

Discuss this entry