January 31, 2012

Comparing VMware View reference architectures

When designing a VMware View environment there are many things you must plan in order to find the right solution. The most important factor in regards to performance seems to be disk io.

During the past few years the server vendors (IBM, HP, Dell) have adapted their own storage systems. Earlier they OEMed much of their offerings and they still do (from NetApp, LSI Egenio (NetApp), dotHill and EMC), but many of the independent storage vendors has been bought by the server vendors during the past few years (3par, LeftHand, Equallogic, Compellent).

On the networking side each of the vendors now have their own 10GbE solutions and most of them are based on Broadcom technology (even Cisco).

Cisco arrived as a server vendor quite late, but has come quite heavily into the market as a serious vendor. They also have their networking offerings, but lack storage. They have however partnered up with both EMC and NetApp so they also have complete offerings. Through the joint company VCE VMware, Cisco and EMC are offering complete solutions including everything you need in software and hardware with single point of support and s single order line.

Each of the vendors also have their own reference architectures for different solution. Last week I was reading through some of NetApp's technical reports about Windows 7 and VMware View. While NetApp's document looks very good on the storage side, it's server recommendation seems a bit old fashioned. Storage may be the most critical factor, but that's not a good excuse for not suggesting an optimal server solution. Memory has become cheaper than it used to be two years ago. A VDI environment using hosts with only 48G memory looks to me like a waste of rack space, power and cooling.  In my mind it's normally more cost effective to have more memory than that in each server.

I thought it would be a good idea to compare the different reference architectures and came up with the following table:

Cisco
VCE
Dell
IBM
VMware
NetApp
HP
desktops/core
10
16
6.5
6.25
12
6.94
5.56
memory/core
32.0
12.0
8.0
10.7
12.0
6.0
8.0
physical memory/desktop
3.2
0.8
1.2
1.7
1.0
0.9
1.4
View version
4.5
4.5
4.5
4.5
4.5
5.0
5.0
Cores
12.0
64.0
156.0
48.0
96.0
720.0
216.0
Memory
384.0
768.0
1,248.0
512.0
1,152.0
4,320.0
1,728.0
VDI hosts
1
8
13
2
12
90
18
Desktops
120
1,024
1,014
300
1,152
5,000
1,200
Infrastructure hosts
0
0
3
2
2
0
0


Each architecture is different and they are also not all using the same versions of VMware View, but  all these reports were released in 2011. Some architectures also describes different environments and I've  not included them all, only the ones that were the most similar. Some vendors have separate servers for supporting infrastructure (vCenter, View Connection server, etc) and some run these on the same servers as the VDI desktops, so it's not a completely apple to apple comparison. All this does however give an idea of what resources that are normally recommended.

In Cisco's document they also increased the number of desktops in their POD (pool of desktops), and cpu was not the limiting factor why it would run slower (not enough memory caused swapping).


As we can tell from these numbers, the average recommended number of desktops per cpu core is 9. NetApp's recommendation is 7. NetApp wasn't actually worst in class in desktops per cpu core as I first thought, but as NetApp is the only vendor suggesting dual quad core servers (8 cores) instead of dual six core servers (12 cores), their solution requires a lot more physical boxes than the competitors. Their scale out design is probably working fine, but looks like a bit overkill and gives some extra costs both on short and long term.

In 2010 NetApp released a 2000 user reference architecture for View 4.0 and here they used servers with 96G ram. 16 servers was here enough to serve 2000 users. Compare that to their 90 servers to serve 5000 users in their latest guide and be confused.



"NetApp and VMware View 5,000-Seat Performance Report"

http://media.netapp.com/documents/tr-3949.pdf


The VMware Reference Architecture for Stateless Virtual Desktops on Local Solid-State Storage with VMware View 4.5 http://www.vmware.com/files/pdf/VMware-View-45-Stateless-RA.pdf


VBLOCK™ SOLUTION FOR VMWARE VIEW 4.5 SOLUTION ARCHITECTURE” http://www.vce.com/pdf/solutions/vce-vmware-view-reference-architecture.pdf


Diskless VDI with Cisco UCS & Atlantis ILIO Eliminating Storage from VDI Architectures http://www.cisco.com/en/US/solutions/collateral/ns340/ns517/ns224/ns836/ns978/optimizing_storage_for_VDI_with_Atlantis_ILIO_and_Cisco_VXI.pdf


“VMware View 4.5 and IBM Storwize V7000 reference architecture” http://www-03.ibm.com/support/techdocs/atsmastr.nsf/5cb5ed706d254a8186256c71006d2e0a/cb4657961a5eb967862578680020fe4e/$FILE/V01_VMware%20View%204.5%20and%20Storwize%20V7000%20Reference%20Architecture.pdf


"Sizing and Best Practices for Deploying VMware View 4.5 on VMware vSphere 4.1 with Dell EqualLogic Storage" http://en.community.dell.com/techcenter/storage/w/wiki/sizing-and-best-practices-for-deploying-vmware-view-vdi-with-equallogic-storage.aspx



"Enterprise Client Virtualization for HP VirtualSystem
Understanding the HP VirtualSystem for Client Virtualization for VMware View 5"

http://h20195.www2.hp.com/v2/GetPDF.aspx/4AA3-8115ENW.pdf

3 comments:

  1. Interesting that Cisco and VCE have different takes on the number of desktops per core, etc., as well as the amount of pRAM per desktop. It just shows that some of these large companies are in that "one hand doesn't know what the other is doing" rut.
    Dave

    ReplyDelete
  2. NetApp Employee here..

    NetApp's architecture that you are referring to is not a reference architecture. It is performance testing report targeted towards analysing the impact of desktops in various scenarios (Boot, Logons, Steady State, 2nd time Logons, Boot after maintenance etc.) on the Storage.

    What was found in the exercise was that Desktop Sizing should not be based on just IOPS/desktop. Also the IO to the desktop is not all Random and it is not 80% Writes and 20% Reads in all scenarios. As per NetApp, there should an assessment done using tools like Liquidware labs to assess the right IO profile. IOPS/desktop is very misleading. More details here https://communities.netapp.com/community/netapp-blogs/virtualization/blog/2011/09/06/netapp-and-vmware-view-5000-seat-performance-report

    BTW, I don't see any link to the NetApp document that you referring to in the blog although you have links to all other.

    You can find a Cisco Validated Design Guide covering VMware View on FlexPod here http://www.cisco.com/en/US/docs/solutions/Enterprise/Data_Center/Virtualization/ucs_view_netapp.html#wp170829

    ReplyDelete
  3. This comment has been removed by a blog administrator.

    ReplyDelete