BM Seer Unofficial thoughts from an anonymous Sun employee

Memory size matters most in watts

Tuesday Jan 08, 2008

Memory size is one of the most important configuration details you need to know when comparing measured server watts. This has been true for a long time.

In 2006 I dug into the actual watts used by 2-socket Woodcrest Xeon servers.

Woodcrest:
330 watts 2-socket 8GB
430 watts 2-socket 16GB
510 watts 2-socket 32GB

Now we have even lower-watt CPUs that have a lot more power management to reduce CPU watts, which means in general that memory size will matter even more in the future.

Another fact is that all of the servers I see using virtualization are being configured with larger and larger memory (32GB to 128GB for 2-socket servers!).

This is why I'm not overly impressed with wattage measured on small-memory configurations like 8GB or 16GB.

[5] Comments
Like this post? del.icio.us | furl | slashdot | technorati | digg
Comments:

So adding 8GB (8GB->16GB) increased power consumption by 100W but adding 16GB (2x as much) increased it only 80W? Sounds inconsistent to me. Have you used the same 2GB FB-DIMMs in all cases and exact same workload?

Posted by I on January 08, 2008 at 10:36 AM PST #

I think you are over looking the effect of having memory act as disk cache. Having 8GB of extra memory in the system may use extra power but may also cause the harddisk to use less power because it does less seeking, and in the future (dare to dream) possibly go into a sleep or low power use mode.

The latests cpu can adjust there clock speed and thus power usage based on work load, having more memory probably allows the cpu to run at a slower speed longer thus further reducing power consuption. If a processor is waiting an extra second or two while data is fetched from the disk instead of disk cache, the cpu could possibly run at a higher clock rate even though it is actually doing little or no work.

Of course in the case of virtualization, having more memory in the box allows it to do more work, and thus replace more machines. And while the memory uses power, Disks and powersupplies and CPU's, fans add up to much more power usage. So I would look at memory power consumption with a lower priority. Even if you have to use 3times the memory in a system to replace 2 boxes, it still turns out to be a win because you removed at least 2 disks, 1-2 cpus, 4 fans, and the cooling requirements of one box in the process... those can easily be 5000 watts of savings per box replaced for the addition of 160watts of power used by the increased memory.

Posted by James Dickens on January 08, 2008 at 11:09 AM PST #

I fully agree that large memory may be required for a server application. I also agree that virtualising applications ito a larger-memory server can save a lot more watts than having lots of small-memory servers.

I don't like vendors focusing their measured wattage demonstrations on very small-memory 8GB, 12GB, or 16GB configurations.

As I've stated before if you really want to save energy the absolute best thing you can do is drive up CPU utilisation and use virtualisation. It saves much more power than the best power management. Do you own measurements and compare the watts/performance ratios.

Posted by BM Seer on January 08, 2008 at 11:32 AM PST #

Not to mention that in the future it will be possible to move workloads to a smaller number of physical machines at times of lower utilization, and then power down the unused physical machines. The flexibility that comes with a fully virtualized environment has the capacity to bring much more in the way of power management in the future.

Posted by Paul Slater on January 08, 2008 at 12:36 PM PST #

I'll go one further, I don't think it is just individual servers, but
racks of servers and sections of a datacenter, or Project Blackboxes - which require less power for cooling than coal-era datacenters :) .

Future virtualisation will also factor in power-cycling and realiability into the transition equations.

Posted by BM Seer on January 09, 2008 at 03:53 PM PST #

Post a Comment:
Comments are closed for this entry.