Uncategorized

Do you want to improve the performance of your AVD hosts for no extra cost?

In my day job, I am lucky to get the chance to do a deep dive into many different customers’ Azure Virtual Desktop environments. For those environments which have been around for a while, and even the more recently deployed ones, I always see people deploying v3 VMs instead of the latest generation v5 SKU’s.

In this blog post, I will perform some testing against the different SKUs and see how much faster they are. As a baseline, we will use the Passmark Performance Test, which runs several tests against the CPU, memory and disk. I will run the test 3 times to ensure it is fair and valid.

SKU differences

But first, let’s discuss the difference between the two different SKUs. The most common SKU I see when reviewing customers’ deployments is the D4S_V3 VMs. The D-Series VMs are known as the Generic SKU, which most people default to when deploying session hosts. The most recent version of that VM SKU is the D4S_v5. So, let’s see the difference between the 2 SKU’s in the table below:

SKUvCPUMemoryTemp Disk SizeMax Uncached disk throughput IOPS/MBPsMax Burst uncached disk throughput IOPS /MBPsMAX NIc/Expected network bandwidth (MBps)Processor type (as tested)CPU Meausred Speed
D4S_v341632 GiB8000/2006400/962/2000Intel Xeon Platinum 8171M 2.60Hgz2.10 GHz
D4S_v541606400/14520000/12002/12500Intel Xeon Platinum 8370c 2.80GHz3.49 GHz

There are two apparent differences between the 2 SKUs.

The processor has changed to a more recent processor. The Intel Xeon Platinum 8171M was released in Q1 2018, whereas the Xeon Platinum 8171M was released in Q2 2019, making it a much more recent processor. The table below highlights the differences between the two processors:

ProcessorArchitectureClockspeedTurbo SpeedCoresThreadsCache Size
Intel Xeon Platinum 8171MSkylake2.6 GHz3.7 GHz2652L1: 1664KB, L2: 26MB, L3: 36MB
Intel Xeon Platinum 8370CIce Lake2.8 Ghz3.5 GHz3264L1: 2560KB, L2: 40MB L3: 49MB

As you can see, the more recent Ice Lake processors have a higher clock speed, more cores and threads, and more cache. This, in theory, will mean a much-improved experience for users using their VDI Desktops on these session hosts.

Testing Results

So, let’s actually see some of the results from the testing. The testing examined CPU, memory and disk capability performance over three separate runs. Below is a table showing the results:

SKUOverall PassMark RatingCPU Mark2D Graphics MarkMemory MarkDisk Mark
D4S_V3 Run 11540.93104.774.31771.71584.5
D4S_v3 Run 21528.42951.184.91736.61581.4
D4S_V3 Run 31520.92895.675.61751.91609.8
D4S_V5 Run 12272.65863.6219.22199.71994.2
D4S_v5 Run 22339.55763.1198.22145.82131.8
D4S_V5 Run 32439.15703.7217.82156.32280.2
D4DS_V52448.25723.6223.52226.02272.4

As we can see, there is a vast improvement in the PassMark score rating for the V5 SKU compared to the V3. This will be due to the increased clock speed and newer processor architecture, with increased cache. What I also found interesting was that the disk and memory were also faster. The D4S_V5 has increased max throughput for the disk, which is why the performance will be higher.

Cost Differences

So, for a vastly improved performance, how much will it cost me to swap from a V3 to a V5 SKU?? Let’s take a look at the cost differences:

10 x D4SV3 VMs @ 260 hours per month will cost you £1087.83 with a 128GB Premium SSD

10 x D4SV5 VMs @ 260 hours per month will cost you £1066.40 with a 128GB Premium SSD

10 x D4DSV5 VMs @ 260 hours per month will cost you £1152.11 with a 128GB Premium SSD

So there you go, it’s actually CHEAPER to run the V5 SKU VMs, but with much faster performance!!

But….

There is one thing you need to keep in mind. With the D4S_V5 VMs, you no longer get a temp disk. A temp disk can also vastly improve the performance of an AVD Session host and is something I always recommend to customers. This is because the pagefile sits on the temp disk automatically, which has vastly improved IOPS (around 5000 vs 500)

So, what is the equivalent SKU with a temp disk?

The equivalent VM with a temp disk is the D4DS_V5, which has a 150GB Temp Disk. I have included the price of that VM above. It is slightly more expensive, but I think it is worth the extra cost. I have included a single run in the performance tests, and it pretty much beat all of the existing scores, so it is undoubtedly the best-performing of all the SKUs.

Summary

For me, swapping from a v3 to a v5 is a no-brainer, and I cannot think of any reason why you would not do it apart from SKU availability in the relevant Azure regions. You get a lot more performance for a cheaper cost.

The only downside is that if you wanted to swap the SKU without rebuilding, you would have to go from a D4S_V3 to a D4DS_v5, which is slightly more expensive but better performing. If you wanted to go from a D4S_V3 to a D4S_v5, which is cheaper, you would need to re-image your VMs or build new hosts.

I hope you found this post helpful, and if you have any questions, please ping me on LinkedIn or reply in the comments!

1 Comment

  1. Great post! I often recommend against Azure SKUs with temp disks simply to eliminate the risk of users saving things there… but I suppose that risk needs to be weighed against the higher performance of a “local” page file.

    I would love to see performance testing using the same methodology for the D4as V5 included as well.

Leave a Reply

Your email address will not be published. Required fields are marked *