AMD No Good For Virtual Desktops? August 23, 2013Posted by audiomatron in Uncategorized.
I wrote some time ago that virtual desktops were not ready for civil engineering firms. I then reversed myself by saying that perhaps they soon will be after seeing that there were going to be advances in graphics hardware for virtual desktop graphics. In fact, this year at VMWorld I see that there are several sessions dedicated to this very topic. Still, at our firm, our physical desktops are still running quite well, and I see no compelling reason to make the move to virtual desktops just yet. However, for the clients of our electronic debris ticketing system, virtual desktops could be quite useful. It is for that very reason that I was tasked with creating a very small View environment so that we could present virtual desktops to our clients, giving them quicker access to their data. Along the way, I discovered something very interesting.
Traditionally, when I build a PC or buy a server, I buy AMD processors. I’ve always found AMD to deliver the most bang for the buck with their CPUs. So naturally, when it came time to build this View environment I sought out a beefy server with AMD processors. This was to be a small setup – 10 desktops on 1 server with local storage. This system would not be mission critical since customer data can also be accessed via web. We picked a Dell PowerEdge R715 With 128GB of RAM and plenty of local 15K SAS disks. This server should have been more than adequate. Our application was to be a Filemaker Pro client, running on Windows 7 64bit, talking to a Filemaker Server running on Windows Server 2008 R2 on VMWare ESXi. Allow me to also mention that we had, for some time, been running a Filemaker client on a Windows 7 VM in our current vSphere environment performing scrips – mostly PDF report exports. As our dataset grew, we started to notice something on this VM that we also noticed on the VMs in our new view environment. Filemaker would either take forever to finish the scripts or choke to death and die, never finishing the script.
My boss was reliant on being able to remote into a machine here in our main office to run these reports, and was becoming increasingly unable to do this on our virtual desktops. It was now my task to determine what was going wrong. In the mean time, I gave him a dedicated, older Dell precision workstation to use for his scripting. This workstation was able to export the PDFs in no time flat. Of course, in my boss’ eyes, virtualization must be to blame, but I wasn’t having it. It was time to conduct some experiments.
As we started to look at the differences and similarities in hardware among the systems in question, one thing stuck out – the workstation has Intel Xeons, the VMs are running on AMD Opterons. I decided to approach my experiments from that angle. Here are the tests I did. Let it be said that I’m never very interested in benchmarks or metrics of that type. I want glaringly obvious, real world proof. Our test dataset here is a 500 page PDF, each page with 3 images and various text.
Experiment 1 – Intel Virtualized
We already had good results with Intel on bare metal. I wiped that old Precision workstation and loaded ESXi on it. It was a quick and dirty install. I created a Windows 7 VM, and installed Filemaker Pro. The PDF exported in just ove two minutes – nearly the same as on bare metal.
Experiment 2 – AMD Bare metal
I installed Filemaker Pro on a production physical server, a Dell PowerEdge 2970. This server is running a resource hog that I like to call Newforma. The 500 page PDF exported in just above two minutes. No problems here
Experiment 3 – AMD Virtual
I actually tried this several different ways – on local storage, on shared storage, 64 bit, 32 bit, alone, with other workloads, on a Dell server, and on a Supermicro server. It didn’t make a difference. In every case, Filemaker would never finish the script and eventually crashed.
Experiment 4 – Filemaker Server on different hardware
We have always run our Filemaker server virtual, but this experiment raised some questions. If the client was this bad virtualized on AMD processors, how could this affect the server? I ran the Filemaker server on a newer physiacl Intel box. There was no change. Everything performed exactly the same as it did virtual on AMD.
Experiment 5 – Intel Virtual on new server
Our conclusion here was that, for our purposes, our virtual desktops running Filemaker Pro need to be on Intel hardware. We bought a new Intel based server for this. The results were consistent with what I saw previously. The PDF exported in just above two minutes.
I think it would be a bit of a stretch to say AMD is no good for VDI or even virtualization. in fact, my experiments and expierence have shown that for server workloads AMD is fine. One thing is clear though – AMD + Virtual + Filemaker Pro = slow. This is only one application, though. Do any of you run virtual desktops on AMD hardware? If so, what has been your expierence?
P.S. – AMD, I still love you guys. We can still be friends, right?