Integrated vSphere enterprise workloads – all together, at scale

One request we’ve been getting is “reference architectures that cover many apps running all at once”.   Asked, heard – and now delivered!

So, here’s another vSphere on EMC reference architecture – large scale, and integrated workloads on V-Max.

• Exchange 2007 - 8,000 user building block, 400MB mailbox, heavy IO profile
• SQL Server 2008 – 75,000 OLTP users against 2 8vCPU, 32GB SQL Server VMs, 1.3TB databases
• Sharepoint – farm of 10 VMs supporting 194,000 heavy users, with a 1% concurrency, a portal configuration using a publishing/collaboration portal and 1 site collection with 15 sites, heavy profile, and total of 1TB of data.

We used Virtual Provisioning to show huge material space savings in the varied use cases.   We used Replication Manager (array-integrated snapshots and clones that also integrate with VMware and Exchange, SQL Server, Sharepoint and other use cases) for rapid test/dev and recovery use cases (including leveraging Kroll tools for mailbox and Sharepoint object recovery)

All at the same time, under load, using DRS, and testing failure scenarioes.  

It showed massive improvements across the board.   As a rule “2x more scale”, and “2x more efficient”.

Interested?   Read on!

Some key takeaways from the testing – there’s LOTS in there, but here are some highlights:

1. In the Exchange test – we used our 4000 user “large” building block (we have standard smaller building block designs for smaller customers) – and vSphere could handle double the number of building blocks compared with our previous VI3.5 testing.
2. In the SQL server tests – we were driving about 20K IOPs per SQL Server VM, each VM supported an average of 650 TPS, and going from 4 vCPU to 8 vCPU configs increased the user concurency by 46% – almost doubling linearly.   We used Virtual LUNs to move the SQL Server databases from one tier to another, live, at the rate of 10GB/min.
3. In the MOSS tests - The document population rate was controlled to around 360 documents per minute, which generated daily data growths of 14.4 GB (600 MB document data per hour).   Using Virtual Provisioning in this use case saved 4.56TB.   We were also able to demonstrate object-level recovery in less than 11 minutes.

 

Here’s the reference architecture doc:

Here’s the Applied Technology Guide:

Comments

[Justin]

It looks like both links for the two documents point to the Reference Architecture.

[Chad Sakac]

Justin - thanks - fixed the link!

Enjoy - and feedback is welcome (including what you would like to see next!!!)

[Rainer]

Hi Chad,
now that the Blog is getting quite large - any chance to getting a search box enabled ?
Rainer

[Mark Burgess]

Hi Chad,

With the huge IO capability of vSphere it is going to be more important than ever that we can size our storage environment for IOPS and bandwidth rather than just capacity.

Are there going to be any improvements with vSphere, with regard to IO capacity planning, that will accurately allow us to size a storage array.

I was hoping that either VMware Capacity Planner would be updated or EMC would provide a tool to analyse the raw data that is generated by the product.

This could then tell us how many flash, FC or SATA disks we will need and what host connectivity would be suitable (i.e. 1Gb/10Gb iSCSI or 4/8Gb FC).

At a minimum it would be useful to have the following:

Peak read and write random IOPS and average block size
Peak sequential IO bandwidth
Filter by time of day (for example so we could exclude out of hours backups)

I believe this is something that EMC has been looking into for some time.

Many thanks
Mark

[Itzik]

Mark,

we DO use the information gathered from VMware Capacity Planner to caluclate how many spindles are needed.

Itzik Reich
Solutions Architect
EMC