Power Readings

These short readings focused on two issues: PUE and CPU advances.

PUE

The problem addressed here is how to begin quantifying data center energy efficiency with a simple metric. The authors propose PUE, defined as (total energy consumption) / (energy consumption of IT equipment). Typical PUEs are approximately 2, meaning there is one Watt of overhead for every Watt that makes it to a server. The authors show that PUE has been going up due to energy prices increasing, and so it is more and more important to use efficient cooling and energy distribution technology.

PUE is certainly a useful metric for determining "infrastructure overhead", especially because this overhead may be easier to decrease than buying new servers. However, another useful metric is "energy per unit work", or even "energy per unit profit". These are the ones that will ultimately determine whether it's more worthwhile to upgrade data center infrastructure, upgrade servers, or just focus on the application instead of worrying about power. The CERN/Intel article looks at some of these issues.

CPU Advances

This article is mainly marketing material from Intel about the virtues of its newer lines of chips. It's interesting because they quote concerns and solutions from CERN. Basically, the power savings benefits of newer Intel chips fall into two categories:

1. Multicore: This development saves power first by putting more CPU into the same physical area (same amount of rack space etc), amortizing the cost of other computer components. Second, decreasing frequencies and adding more cores improves performance-per-Watt because performance drops linearly with frequency but power drops quadratically (or cubically?). This is a very interesting effect that's actually also being exploited in cell phones and in radical computer designs like CMU's FAWN. It will be interesting to see just how far this might be pushed - will some cloud applications run on ARM nodes, or will the overhead of more cores just be too much to make this worthwhile?
2. Virtualization: This technology saves power by enabling consolidation of low-utilization machines. However, the article rightly points out that virtualization may not make sense for resource-intensive applications because it will just add overhead and increase power cost. This technology is probably not as important to CERN as multicore, but might be important to many enterprises who have a bunch of low-utilization boxes lying around.

Finally, one interesting point in the article was the constraint of floor space and data center building design that CERN was facing. Organizations can't just move to a new data center without incurring high costs. Multicore and virtualization both save space, not just power, so that was one of the other reasons CERN was looking at them. This is one area in which cloud providers may have an advantage, by buying cheap land and designing data center buildings to enable the fancier cooling technologies that were not possible at CERN.