Suitable GRID

Hi all.

I am a new member on the forum, nice to meet you guys

I have a question or needing a few advice from advanced individuals on the forum.

So, at my work, we are adding in 2 more esxi in order to expand the use of VDI.
We already have 70 users who use VDI (sales department) who use a thinclient (horizon 7) and PCoIP zero client.

We are planning to add in around 100 more users on the two Esxi we are putting in. But this, time it’s for users who work in the production department

These new users won’t be doing heavyload editing work but it’s more for reviewing edited video files and clips.
We are planning to hand out the thin client with blast extreme (horizon 7.1)

My question was in order for them to view and playback 1080p clips and videos smoothly and with ease, will a M10 be good enough? We are testing with K2 right now and apparently my co workers seems to be happy with it. I wanted to know if spending almost double the money for M60 was worth it? By the way out of the 100 users, maybe not even half of them will be actually watching clips simultaneously.

If anyone could give me an advise, i would appreicate it very much.
TLDR;

  • adding 100 more VDI users for work
  • half of them will be viewing 1080p videos often
  • will a M10 do the job as of smoothness and playback ease

Thank you.

Hi

Welcome to the forum!

Diving strait in … You’re going to use Blast Extreme, which means you’re going to use NVEnc. NVenc requires a minimum of a 1GB vGPU profile. Based on that, each M10 will support 32x user’s maximum, so you will need 2x M10s per server to cope with that amount of users.

The 2x new servers that you’re adding, what specification are they? RAM? CPU? (CPU model number and generation (v3 / v4))

Which Windows Operating System are you going to be delivering?

As I’m sure you’re aware, when you move to Maxwell, you move to a CCU licensing model, so don’t forget to factor this in to your costs. With the information available, it’s looking like you will want to purchase vPC licenses.

Your question:

Q - Is the M60 worth double the amount of the K2?

A - With the M60, you get more performance and double the Frame Buffer (2x 8GB, instead of 2x 4GB). Therefore, based on Frame Buffer allocation, you can get twice the amount of users in the same server chassis. So, for K2s to support the same amount of users as a pair of M60s, you now need to purchase 2x additional K2s, plus another server to put them in, plus you need rack space to install it, power to run it, support and maintenance for the server, another Hypervisor license and lastly, Air Conditioning capacity to cool it. All of a sudden, that increase in price isn’t looking so bad ;-)

For example, I have a few high performance clients who are space limited in their datacentres (they physically have no space left in their racks). So space (user density) is a real premium to them, and the increased price of the newer GPUs is insignificant compared to what it would cost to scale into another datacentre, and for some of them, there isn’t even that option.

So when comparing GPU costs, try not to look at it as simplistically as comparing old to new. Look at what it costs to support "X" amount of users in a given set of circumstances, then you’ll find the true value of the newer GPUs. And regardless of what the GPUs are purchased for (in terms of a monetary value), the real value of each GPU will vary per customer depending on their circumstances.

Hope that helps

Regards

Ben

Thank you for the quick responce!

We are still in the deciding phase of what server and what kind of hardware we are going to hand out to the user.
But as to OS i think we will be deplying Windows 10 on the thinclients (maybe i5)

To be really honest with you, we dont even know the exact number of users we are going to be deploying to but its around 50 to 100 mark.

I understoon your theory about M60, but since the sole purpose of this project is to let the users view edited clips and videos, not actually editing them, will a M10 be more suitable? Since partial users will be using it for video viewding purpose and others will be using it for normla tasks (deskwork apps etc)

We are only testing it with the K2 that we were able to borrow from elsewhere, we are trying to put in a newer GPU hence we are deciding on an M10 or the M60, my understanding is that M10 is more user density friendly with mid level 1080p decode/encode and video viewing purpose, where as the M60 is the opposite.

Also, what type of integrated GPU will be suitable in order to decode the 1080p file on the thinclient side?

I am sorry for my lack of knowledge since i was put on the team of the project not so long ago and i only started to study and search about this area of computers 2 weeks ago.

Once again thanks Bjones and i would like to hear more and if anyone else on this forum can give me anything that would mean alot to me!

If you’re delivering Windows 10 as your VM OS, then you need nothing less than 1GB of FrameBuffer

As your user count and density are directly tied to your vGPU profiles, you do need an accurate count of how many users you plan to move on to the system. Otherwise, you will either under-spec everything, meaning that you will need to purchase more individual servers as they won’t be a sufficient specification, or you will over spec them, and have wasted resources. You do need to be more accurate with your numbers than just saying “around 50 to 100”. Also, what about the rest of your infrastructure? Have you considered the additional storage I/O and network bandwidth? … when adding 50 - 100 users who are working with video, you need to consider all resources.

It’s not a theory, it’s a fact :-) Don’t look at individual component prices, look at the cost of the whole system. The more users that can share the same resource, the more cost effective the entire system becomes, and that’s one of the main reasons we virtualise (and the ENTIRE reason that NVIDIA GRID even exists!); to get better resource utilisation, ultimately, because it ends up being cheaper overall than giving users their own dedicated hardware. It’s not just about the physical server; Rack Space, Power, Cooling, Licensing, Support, Maintenance, Upgrades, Updates etc etc. It all costs money.

The higher the server density, the more cost effective the system is to implement and support.

You didn’t ask for a performance comparison between the M10 and M60, you asked whether the M60 was worth twice the purchase price of a K2, to which the answer was, (in a little more detail) there’s more to it than unit purchase price. The M60 is obviously more powerful than the M10, however the M10 trades performance for density. Based on FrameBuffer, the M60 will support 50% less users than the M10. This is where you need to know how many users you plan to add to the system, so you can spec and design appropriately.

Let’s say you have 64 users total (that falls within the “around 50 - 100” bracket):

If you have up to 32 users per server, I would say buy 2x M60s for each server and give them a 1GB vGPU profile each. The reason for that, is that the M60 has more H.264 streams than the M10 (36 vs 28), and as you plan to deliver video, this is an important consideration. If you want to add more users on to the system, you now need to factor in what I mentioned above about cost vs value. You now need to purchase additional server resource to support the additional user density. Also, if you have less users on a server, you need less CPU and RAM to support them; this is money that could be used to offset additional server resource to support more users.

Let’s say you have 65+ users total (that exceeds the M60s FrameBuffer capacity with 2 servers):

For more than 32 users per server with a 1GB vGPU profile, you will need to use the M10. The M10 will do the job, however if all users on the same server start viewing videos at the same time, you may well start to suffer reduced performance. Remembering that once users are connected, you can’t load balance VMs between hosts (however you will be able to later this year / next year …).

Also, your server density will depend on your CPU and RAM specifications, which you haven’t mentioned yet.

With all of that said, if you only have 2x new servers and want to support more than 64 users, a pair of M10s in each server is your only option, as the M60s simply don’t have the FrameBuffer capacity to support more than 64 users.

You need a Thin Client that has H.264 decode capability. Typically something with a SOC (System on Chip) or dedicated GPU will have that capability. A “Raspberry Pi” will be the cheapest option, then you’re in to many different choices. Intel NUCs make great Thin Clients, and there are a lot of different models to choose from and if you don’t need the latest and most powerful, they are also pretty cheap as well. The Thin Client is still a very important part of the system. Poor Thin Client choice can destroy a perfectly designed back end system due to the user experience, so make sure you test various vendor brands / models and validate them properly. And don’t underestimate what’s required to stream a video to an acceptable level. As I’ve mentioned to someone else who posted on here recently, I personally think streaming videos is one of the hardest things to do, as it requires near perfect visuals and audio at all times. Any drops in visual quality or frame rate are immediately visible to the person viewing the video, where as someone using Outlook or typing a document, may not notice a drop in visual or audio quality.

Regards

Ben

I read that M10 can support up to 64 per card, 128 per server. After talking to my co-workers it turns out the number of users we are deploying will be more than 60 and less than 70. So as you mentioned below, 2xM10 will be our only choice regarding this project. As far as other resources, there is other people working on those department and it seems we have a go, plus we are using a pretty heavyload Netapp storage so as to storage goes i think we are ok.

Sorry man! My english isnt the best and used "theory" to sound cool i guess lol, I completely agreee with what you said about cost and value. I guess it all depeneds on how much money they gunna let us use for this project. But so far as to cost and value, like you said once again, M10 wins out maybe since M60x2 can only support 64 users.

Once again i apologize for the misinformation, Althoguht i said a thinclient, the hardware we are trying to deploy is the notebook type (thinkpad) and i have read that in order for the thinclient to decode the h.264 files it would need a integrated GPU above Haswell for horizon 5.x and onwards? So by saying that, if we were to get a thinkpad for the purpose of our agenda, will skylake i5 will be a wasted resource like you mentionded above? Also we will be using Win10 IOT edition or whatever on a Windows embedded inside the thinkpad… Oh and once we have decided on the server, i will sure to post the system spec!

Once again thank you so much man for replying and hope to get a reply from you soon

Regards

kaeon

Forget about the M10 supporting 64 users per card. Yes, “technically” it will based on FrameBuffer, so factually that is correct. However, as a 512MB vGPU profile (that’s what you have to use to get that density) is not enough to support a modern Operating System plus applications, you need to halve those numbers as you need nothing less than 1GB of Frame Buffer, which will give you 32 users per M10. Also, as mentioned above, to use NVenc, you need to use a 1GB vGPU profile, again, halving the 64 user number.

That’s still not a definitive answer regarding user density :-) And between 60 - 70 users makes it even more important to be absolutely accurate, as 64 is the cut off point from one model of GPU to the other. A pair of M60s (2 in each server) will support 64 users across the 2 servers using the same vGPU profile that you will be using on an M10 (1GB). However, as mentioned, you’ll have more H.264 streams available on the M60s. You need to give a definitive answer on user density before you choose your GPU model. If you have 65 or more users, then the M10 is your only option, but up to and including 64, the M60 is still an option.

All these numbers are based on a 1GB vGPU profile (which is still very small). Have you done any monitoring yet to see how much FrameBuffer you actually need? (Again, with 1GB being the minimum). If you need more than 1GB of Frame Buffer, then this obviously changes things …

Skylake is a good CPU, however as always, you’ll need to be specific on the CPU model number as they will have different GPU capabilities. If you’re buying new, Kaby Lake is better …

So we can see it, specifically, which Thinkpad model are you looking at? Do you have a URL? …

Regards

Ben

Oh man im sorry, yes youre completely right lol, i wasnt thinking straight, so yes! 64 per server :)

From what i just heard from my co-worker, we are planning to put in 2 servers and 1 card each. (from finalcial perspective, buying 4 M60 might just be impossible lol)

About the "loadbalancing being available sometime in the near future" do you mean we will be able to use vMotion with the vGPU?

But we have a few thinclients we are looking at. I will post the link before (the ones we are looking have AMD cpus)

http://h20195.www2.hp.com/v2/GetDocument.aspx?docname=4AA6-1284EEE&doctype=data%20sheet&doclang=EN_GB&searchquery=&cc=uk&lc=en
or
http://www8.hp.com/h20195/v2/GetDocument.aspx?docname=4AA6-8039ENUC

Thank you for the quick responce once again.

Regards

kaeon

Ok. That won’t work. You won’t have enough GPU resource.

64 users per server = 128 users total (that’s completely different to anything you’ve posted above(?!))

You WILL need 2x M10 per server. Anything less, and you won’t hit those numbers.

Unless you have 2x M10 in each server, you won’t be able to load balance anything, as you won’t have enough resource.

You can’t guess this. You need to be accurate with your numbers, calculations and specifications or the project will be an expensive failure. And nobody wants to see that.

Regards

no no, what i meant was, the total amount of users we CAN support is max 128. But in reality, its 68 (i got this number just before) and we i said 1 card each but i meant 2, so if we end up needing to add in more users, we could

After talking to other departments, we came to the conlcusion that not all 68 user will be watching videos simoultanouesly, let alone maybe even 56. But for future consideration, we a system thats user desnsity friendly.

Sorry for mmaking you confused man… i apologize.

Ok, this is getting confusing now.

You have 2 servers and each server has 1x M10. You have 64 users in total. 1 server will have 64 users on it. The other server will sit there doing nothing unless there is a failure. Is that correct? If yes, that won’t work.

With what you’re trying to do, to support 64 users on a single server, you WILL need 2x M10s for the reasons mentioned above about “NVenc” (if you’re not sure what “NVenc” is, you need to read up on it as it is very important to your solution: https://en.wikipedia.org/wiki/Nvidia_NVENC ) and Frame Buffer allocation.

Currently, you cannot vMotion VMs with a vGPU assigned. In fact, VMware has made no mention of introducing this feature. The only vendor to announce that is Citrix, however it would seem logical that VMware are also working on this solution, but currently, there is no sign of it.

Regards

Ok here is our plan.

we will have 2 esxi servers.
we will put in 2 M10 card each on these servers.
we have 68 users in total.
our plan is to put 34 users in each server and they will all be using the Blast Extreme protocol via horizon 7.1

Im sorry if im confusing you but this is the simplest way i can put it with my knowledge :(

Regards.

That’s closer, but still not right.

Based on that user count (68), if you need to perform any updates or maintenance that require all users to be on a single server, or one of your servers fail, you will not be able to support all 68 users.

Each M10 will support 32 users with a 1GB vGPU profile. So with 2 M10s in each server, you can support an absolute maximum of 64 users per server.

Page 5 & 6 of this guide will give you the vGPU profiles of the M10 and tell you how many users you can get per card: http://images.nvidia.com/content/grid/pdf/GRID-vGPU-User-Guide.pdf

Regards

i see… thank you so much for going in detail man

i apologize once again for confusing you.

i have told my co workers and we decided to cut down the user count to 50 max.

so from what you say, we should be ok…!

im still yet to decide on the thinkpad though lol -.-

Either laptop will be ok, as they have an AMD APU. However …

Do some comparison searches on the difference between the AMD A8 PRO-8600B APU and the AMD A8 PRO-9600B APU to work out which is better suited for your use case.

You’ll find that one is clocked higher but the other has more memory bandwidth. Up to you on which to choose.

Regards

Thank you very much BJones, you helped me out alot.
Today we receive the mt43 as a test drive and it seems to work flawlessly.

We will deply 5 mt43 to users for a test run and see what they say.
We will still be using the K2 for the testing phase.

Once again, thank you mate.