Tuesday, November 24, 2009

Server power

Many IT shops are looking to reduce power and the current solution is virtualization and moving toward local clouds. Little effort is put into the servers in the data center. Now certainly virtualizing 10 servers into 2 physical systems is likely to give a good deal of power reduction. However there are still the 2 physical servers, along with other servers that may not be appropriate for virtualization, and possibly a computational GRID system.
Is there much savings in the servers themselves? of course, and possibly quite a bit. There are two things to consider here, internal components (cpu, ram, disk, power supply, etc) and the chassis itself (2 cpu system, 4 cpu system, etc.). For chassis itself this is often dictated by the requirements. Try to avoid purchasing a system with future expansion capabilities. If 2 cpus are required, purchase a system that can handle just two cpus. You will waist a good deal of energy for the expansion capability, and worse by the time you are ready to expand the system, Moore's law generally means that it is cheaper to purchase a new one than to expand.

For the internal components though, you find different situations. It is well known to look at the CPU for power differences. Most CPUs have known power draws, though the draw at any given time is based on load, and modern CPUs have many power savings features for when there is not a heavy load. Server CPUs often vary from 45 to 120 watts, however performance can vary with this power consumption. Sometimes the choice can be pretty easy to judge. Say a server is needed with the performance capabilities of two Nehalem CPUs at 2.26 GHZ. This could be met by either two E5520 CPUs at 160 watts or two L5520 CPUs at 120 watts. The L5520 CPUs will likely cost a little more but with almost 1KW savings per day with the same performance.
On a basic server system today, server processors consume the most power, followed by memory, then disks or PCI slots, the motherboard, and lastly the fan and networking interconnects.

For example say specifications require a dual cpu system with 24G RAM. With a Nehalem architecture the RAM could be configured at least 3 different ways; 12x2, 6x4, and 12x2. Further the RAM can be a mix UDIMMs and RDIMMs as well as single, dual, and quad ranked. These configurations are hard enough to work out just to get a working system but there can be a huge difference in the power consumption based on the RAM configured, worse even the manufacture of the RAM. It is pretty easy to configure two seemingly identical systems of the same make and model with different RAM configurations such that one draws nearly twice the power of the other. Some server manufacturers give some web based tools to give an indication of how much power is required for a given configuration. Generally more ranks on the DIMM the more power efficient, also use higher density DIMMs for overall system.

This savings will be slightly more than doubled at the meter. For every watt saved in the computer, the average data center saves an additional:
  • 0.04W in power distribution
  • 0.14W in UPS
  • 1.07W in cooling
  • 0.10W in building transformer and switch gear
The savings could be much more significant than just the electric cost. It is common to fit 40 1U servers or 64 blade servers to a rack. Even with just modest changes a savings of 100 watts per server is possible, which is easily achieved with combination of CPU and RAM arrangement. With 64 blades per rack this comes to 6.4KW. This could be the difference between the infrastructure being able to support the servers and needing to upgrade the power distribution, UPS, and cooling in the data center.

The hardware used for virtualization, local clouds, and compute intensive environments are taking an increasingly large portion of the data center power budget. The configuration of servers for these systems have to balance initial cost, licensing, support and now power use. The SPECpower_ssj2008 benchmark can be used to assist though the systems are often not optimally configured for virtualization or GRID work. What is needed is a merger of SPECpower and the (currently unfinished) SPECvirtualization and/or SPECmpi2007.
Anandtech has an excellent article on server performance to power.

No comments:

Post a Comment