So i'm building out this Virtual infastructure, (ESX) and i'm trying to make intelligent decisions related to acceptable sizes of servers.

16 (15 usable) The hosts that i'm running on are 8 Core dual socket with 256 GB ram..

I'm trying to avoid crossing boundraries related to NUMA (8cores 128GB ram)...

So maximum server size i can build is an 8 core with 128GB ram

now the problem is I'm trying to offer flexability in what I offer workload wise, but still maintain enough control that I don't build stupid things..

For simplicy sake, lets assume I can't oversell any of the resources so a phyiscal core is = 1 virtual cpu

I'll give you an example of a stupid configuration:

group of scientists A want a 2 high performance computer, but they don't require much memory.
So i build them an 8 core machine with 16 GB of memory. I can only put 2 of them per host.

Each one of these computers eats up an entire node and i end up throwing away 212GB of memory per host. because I don't have any CPU left.

Now I know i'm going to have to support some loads like this, but if I make them smaller say 4 core with 8 GB of memory and build 4 of them. I give myself some flexibility.

I could spread the 4 across 2 hosts, and hope that I have another load that has low CPU overhead, but high Memory Overhead.

However at some point I hit a point of diminishing returns because all servers have some overhead not related delivery (running the opearting system, nic antivirus et cetra). In addition the management/patch/update footprint gets bigger.

I'm trying to figure out the right size building blocks that we can offer up to allow us to reduce overall capacity waste. If we follow prior patterns and let the customers choose what they want, ill end up with 8 CPU machines with 4 GB of memory like above which will eat up hosts quickly, and not allow us to provide additional services to more balanced workloads.

A balanced workload above would be something like a 8 CPU with 128GB ram server or 2 4CPU 64GB ram servers.

Anyway, anybody got any thoughts or ideas on how to demonstrate the packing efficeny and how to visually demonstrate waste of unreasonably non balanced shapes?

[Edited on April 9, 2015 at 7:40 PM. Reason : dd]

4/9/2015 7:38:36 PM

A pratical example related to above.

I have terminal services machine, on which users do office productivity work. They run regular office apps, but all of them love opening every office application and leaving them open all day.

So low CPU load, but high Memory overhead.

It may be better to spread the actual apps across multiple servers that are more idea in shape than to have a server with 4 cores and 128GB of ram, because at least that way i'm not stressing the cpu, and ESX doesn't have to move things around as much.

4/9/2015 7:56:30 PM

A pratical example related to above.

I have terminal services machine, on which users do office productivity work. They run regular office apps, but all of them love opening every office application and leaving them open all day.

So low CPU load, but high Memory overhead.

It may be better to spread the actual apps across multiple servers that are more idea in shape than to have a server with 4 cores and 128GB of ram, because at least that way i'm not stressing the cpu, and ESX doesn't have to move things around as much.

4/9/2015 7:56:30 PM

A pratical example related to above.

I have terminal services machine, on which users do office productivity work. They run regular office apps, but all of them love opening every office application and leaving them open all day.

So low CPU load, but high Memory overhead.

It may be better to spread the actual apps across multiple servers that are more idea in shape than to have a server with 4 cores and 128GB of ram, because at least that way i'm not stressing the cpu, and ESX doesn't have to move things around as much.

4/9/2015 7:56:30 PM

format c:

4/9/2015 7:59:54 PM

I read this as "Parking/capacity planning problem" and expected a thread about an overcrowded parking lot.

4/9/2015 9:43:15 PM

you have 2 options I see.
1 - forget vmware and just use docker and deploy your workloads as there's capacity. the docker manifest will handle hardware partitioning
2 - build as many of each config type will fit on a host or in a cluster.

ie: 2x 64x8
4x 32x4
8x 16x2

then keep track of which are in use on the hardware at any given time and only allow what you have capacity for to boot at any given time

you can do this on the fly with vmware's api, which sucks at large scale, but should be fine on this scale.

the idea of permanent hardware partitions is mostly antithetical to the benefits of virtualization

[Edited on April 10, 2015 at 6:05 AM. Reason : this assumes an on-demand access model, not vdi. with vdi just slice it in lowest common denominator]

4/10/2015 6:03:27 AM

we have a mix of XenAPP/XenDesktop / VDI... Right now VDI is the easier item to handle as they are pretty cookie cutter.

Right now we don't have to really worry too much about capacity, but i want to ensure we don't start building things that aren't supportable long term, with the hardware that we have.

I'm trying to ensure that leadership understands the diference between 500 people using outlook, and 500 people running ArcGIS, and that we can't sell both use cases at the same rate.

[Edited on April 10, 2015 at 7:55 PM. Reason : dd]

4/10/2015 7:48:59 PM

if you have xenapp, why not just virtualize the apps that aren't in the lowest common denominator desktop and publish them to the users that need them and give everyone the cookie cutter? you get better usage and can buy your software for concurrent user count instead of per seat

4/10/2015 10:10:07 PM

buy more vmware please

4/10/2015 11:25:39 PM

^^

That is the current plan. XenApp does let us balance out a load on a single server (i.e. high memory app vs high cpu app).


I'm hoping to have something in the summer put together so we can do some grid/3d stuff, and frankly most of this is planning for a year or 2 down the road... But if we are going to be successful i think you need a 2 year plan, not a 2 month plan.

4/13/2015 8:38:10 AM

I would highly encourage you to play with citrix and grid in amazon on a g2. You might be disappointed

4/13/2015 12:36:03 PM