Power

The Opteron may have a slight disadvantage here as it has a 1200W PSU instead of a 1000W unit. We had no choice as our 1U server with 1000W PSU that we used in our previous reviews is not supported by ESX. Our Supermicro SC828TQ-R1200LPB 2U works very well with ESX, but it may consume a few extra Watts. (Or perhaps not - without testing both PSUs we really don't know their efficiency curves.) All Xeons use the exact same chassis, motherboard and power supply. We decided to test with both 16GB and 32GB. In our benchmarking scenario we only needed 16GB, as our VMs were very processing intensive and didn't need more memory. In the real world, it is much more likely that most of your VMs need far less processing power and will require more memory (especially when you run more VMs). The extra data point allows you to understand how much power eight extra DIMMs require, allowing you to estimate how much power your setup will consume.

System Power Comparison
  Typical CPU load Power with 6 VMs
at full load (16GB)
Power with 6 VMs
at full load (32GB)
Xeon 7330 (2.4GHz) 91-96% 456 541
Xeon 7350 (2.93GHz) 90-95% 605 692
Xeon X7460 (2.66GHz) 85-90% 502 587
Opteron 8356 (2.3GHz) 95-99% 485 534

It's amazing to see that the 6-core X7460 is capable of staying well below its older 65nm brother, the X7350 at 2.93GHz. Both CPUs have the same TDP, but the 6-core X7460 consumes 25W less per CPU than the X7350. Part of this can be explained by slightly lower CPU usage as the 24 cores do not have to work as hard to handle the six VMs. The X7460 is a huge chip, but that does not prevent the newest Xeon from performing 45% better while consuming 25W less per CPU.

Transactions per Watt
Xeon 7330 (2.4GHz) 3.93
Xeon 7350 (2.93GHz) 3.26
Xeon X7460 (2.66GHz) 5.59
Opteron 8356 (2.3GHz) 5.26

This table also helps explain why it is so important for Intel not to wait for a Xeon Nehalem MP to regain the performance crown. The current Xeon 73xx MP line takes a serious beating from the Opteron when it comes to virtualized performance/Watt. However, this table also makes it clear that the current Opteron has no chance of beating the 45nm Xeon MP. If we use the 8360SE (2.5GHz) instead of the 8356, performance will rise 8% at the most, but power consumption will probably increase by 20-30%, resulting in an even worse performance/Watt result.

ESX 3.5 Update 2 Virtualization Results Limitations and Conclusion
Comments Locked

34 Comments

View All Comments

  • JarredWalton - Tuesday, September 23, 2008 - link

    Heh... that's why I love the current IBM commercials.

    "How much will this save us?"
    "It will reduce our power bills by up to 40%."
    "How much did we spend on power?"
    "Millions."
    [Cue happy music....]

    What they neglect to tell you is that in order to achieve the millions of dollars in energy savings, you'll need to spend billions on hardware upgrades first. They also don't tell you whether the new servers are even faster (it's presumed, but that may not be true). Even if your AC costs double the power bills for a server, you're still only looking at something like $800 per year per server, and the server upgrades cost about 20 times as much every three to five years.

    Now, if reduced power requirements on new servers mean you can fit more into your current datacenter, thus avoiding costly expansion or remodeling, that can be a real benefit. There are certainly companies that look at density as the primary consideration. There's a lot more to it than just performance, power, and price. (Support and service comes to mind....)
  • Loknar - Wednesday, September 24, 2008 - link

    Not sure what you mean: "reduced power requirements means you can fit more into your DC". You can fill your slots regardless of power, unless I'm missing something.

    Anyway I agree that power requirement is the last thing we consider when populating our servers. It's good to save the environment, that's all. I don't know about other companies, but for critical servers, we buy the most performing systems, with complete disregard of the price and power consumption; because the cost of DC rental, operation (say, a technician earns more than 2000$ per year, right?) and benefits of performance will outweigh everything. So we're so happy AMD and Intel have such a fruitful competition. (And any respectable IT company is not fooled by IBM's commercial! We only buy OEM (Dell in my case) for their fast 24-hour replacement part service and worry free feeling).
  • JarredWalton - Wednesday, September 24, 2008 - link

    I mean that if your DC has a total power and cooling capacity of say 100,000W, you can "only" fit 2000 500W servers in there, or you could fit 4000 250W servers. If you're renting rack space, this isn't a concern - it's only a concern for the owners of the data center itself.

    I worked at a DC for a while for a huge corporation, and I often laughed (or cried) at some of their decisions. At one point the head IT people put in 20 new servers. Why? Because they wanted to! Two of those went into production after a couple months, and the remainder sat around waiting to be used - plugged in, using power, but doing no actual processing of any data. (They had to use up the budget, naturally. Never mind that the techs working at the DC only got a 3% raise and were earning less than $18 per hour; let's go spend $500K on new servers that we don't need!)

Log in

Don't have an account? Sign up now