VMware ESXi 2U Server Vs. Traditional Bare Metal 1U Server

Question

I have a Dell PE 2970 quoted and ready to order with the following specs:

• 2X 6-core AMD 2.2ghz 6mb HT-3
• 6X 73gb 15K SCSI 6Gbps 2.5in (RAID 10)
• 16GB (4X4gb) 800mhz RAM
• 2X Intel Pro 1000PT Single Port 1gb NICs

How will the above server perform compared to traditional bare metal server I am running currently? Current server is a 5 year old PE 1425, 2X 160GB SATA, 2X Intel 2.8ghz single-core, 2GB RAM.

Primary load is LAMP web server traffic, not much, max 50GBs per month.

I assume the PE 2970 running ESXi will handle the "load" with absolute ease, but I'm planning on adding development environment VMs for Java/Grails & Ruby on Rails (both CentOS 5.5) and Windows 2003 Server, which will all be routed via one of the NICs (primary NIC will be dedicated to CentOS 5.5 LAMP VM).

Note sure where the bottleneck will be performance-wise in the PE 2970, but trying to account for issues now before I take the plunge. I'm buying a Cisco ASA 5505 for the firewall as well.

Suggestions appreciated!

Answer 1

The processors give you about 6 times the CPU grunt of the older Intel, RAM has (I think) about 2-3 times the bandwidth of the older Intel, Hard disk subsystem is about 9x the IOPS 6x the throughput of the older system. No real surprise there, I'm sure you'd figured that out having selected the components.

In typical environments I'd be looking to consolidate 6-8 "average" servers similar to the baremetal one you describe onto one of these when virtualizing. The 2970 is a pretty good mid range server and the 6 Core AMD Instanbul Opteron CPUs are quite good at virtualization as they support NPT/RVI.

However I'd call out a few things to consider.

1. Dell are about to release the R715 (the reference is buried in the middle of this press release) - this is their 11G successor to the 2970 which is based on their 9G platform. The R715 supports the Magny Cours 6100 series Opterons that are substantially better than the Istanbul CPU's and have a whole lot more cores (8 or 12 depending on the model) and memory bandwidth. The other platform improvements in the 11G line make it quite appealing but as with everything shiny and new it will cost more.
2. As you consolidate more stuff into a single box you increase the impact of any outages or downtime. Every extra server VM you add makes it harder to shut the damn thing down for maintenance. That's fine so long as you keep an eye on it but dealing with that is one of the reasons people pay so much to have shared storage and VMotion capable licensed VMware setups.

Answer 2

The new hardware should perform better than the old stuff. Your disk I/O should be markedly better since you'll have more platters involved, and they'll be rotating faster. The environment you describe doesn't look terribly disk I/O bound as it is, so you're more likely to be constrained by memory. At first the environment should be markedly faster than the old one. As you grow, you may end up with some CPU problems as runaway dev-processes chew resources, but that can be limited by not assigning every VM its max VCPUs. Looks solid to me.

Answer 3

My rule of thumb is that you can consolidate at about 2 real machines to 1 core that's virtualized as long as the services you are migrating aren't super CPU bound. With a 12 core host you can probably put 24 VMs on it comfortably as long as you have the RAM and IO. You're not going to be able to do better than a 6 disk RAID 10 in that chassis so that looks good. If you are allocating 8GB of disk per VM for a default install then that would be about 192GB total which should be pretty close to how much space you actually have.

If you allocate 1G of RAM to each VM you can easily run out of 16G of RAM. If you can afford it and if you think that you will end up creating enough VMs to need it, then try and get 32G instead. I think that will be better sized to the capabilities of the host so that there is no particular bottleneck, you'll run out of disk, cpu and ram all at the same time. Better to get it now than try and put it in later, you may not want the risk and downtime of fiddling with the hardware or you may not even be able to get RAM for that kind of host at a reasonable price down the road.

Of course if you just don't need it then what you've built should otherwise be fine.