Sunpyo Hong and Hyesoon Kim focus their article upon energy conservation in GPUs by analyzing the optimal number of active cores vs. power consumption. Most people would probably muster an uninformed guess that power consumption for GPUs (or any computer model, for that matter) increases and decreases on a linear scale. In other words, one might assume the best ratio of power consumption and performance would be during a “power off” or idle state, or at max consumption/capacity. Unfortunately, such is not the case, and Hung/Kim can, by pinpointing the optimal active cores, “save up to 22.09% of run-time GPU energy consumption and on average 10.99% of that for the five memory bandwidth-limited benchmarks.” To do so, they obtain power model parameters to stress the different component of the GPU's architecture. Modeling different areas such as temperature and consumption by analyzing different inputs such as memory, they create an IPP (Integrated Power and Performance) Model, which indicates the performance and energy consumption of different components. Based on run-time statistics and graphs, the IPP predicts GPU power consumption, as well as the previously mentioned optimal active running cores, and performance per watt. As a result, the findings in this article and the proposed IPP can be used to optimize configurations for programs to effectively (and efficiently) use GPUs in the most productive/conservative ratios.
Sunpyo Hong, Hyesoon Kim, "An integrated GPU power and performance model", ISCA 2010.
---
Prior to GPUs being a hot topic for power consumption in computers, extensive research was done on CPU power consumption. The majority of what I learned through this article was really the basic, systematic approach to a proper power model and the components that comprise that model. Some of these rules of thumb include accuracy, generality, speed, and expense. The authors used 5 different systems, all of varying degree in power/performance potential, along with various modeling tools, to evaluate each system. From these models, a noteworthy item I came across was that the disk benchmarking for ClamAV was the most accurate because of the minor power consumption contribution disks have vs. CPUs. I considered it noteworthy because the article later mentions that there is a lack of insight into disk and memory power consumption even though CPU power consumption is decreasing, which is limiting the accuracy of benchmarking memory/disk intensive systems (Pg. 4). Our leader for the CREU project, Dr. Rivoire, also mentioned to us this week that with models that aimed to stress the disk's power consumption (while running ClamAV), there was a large variation in power consumption as opposed to the small variation on a normal system, and the subsequent results were anything but accurate; which brings the idea of ClamAV being the "most accurate" analysis method of disk power consumption into question. I also found it interesting, and would be curious to see, how the cooling system of a GPU might be factored into the equation, which is mentioned under the conclusion section. It is such because the cooling system would also consume power, and it would be interesting to see the complexity of the code that would go into making sure that component is reported to the OS (as also stated in the conclusion).
Suzanne Rivoire, Parthasarathy Ranganathan, Christos Kozyrakis. A Comparison of High-Level Full-System Power Models. Hotpower, 2008.
------
It was interesting to compare this article with having experienced growing up in the 90's and early 21st century. PC gaming, software development, and GPUs in general developed an incredible amount over about a decade, and it is amazing to see where everything lies now after this article highlights a few of these points. One thing that was clarified for me was that I had a vague idea of the software/GPU relationship, naturally, because of gaming. But, I didn't necessarily grasp it as much as I do now after having read the article. And, although the concept of a GPU seems quite complex now as a CPU did 10-15 years ago for the average person, I believe that the knowledge (and expectation) of GPU specifications is only going to become mainstream once software is streamlined to take full advantage of them. The article mentions at the end that if you're “Twittering or checking email, GPGPU computing is not for you”, but I must disagree, in part. In due time, people will expect the same technology and graphics of a Playstation 3 or high-end Nvidia graphics processor to be emulated on their daily use PC, or better yet, their cellular phone - Budge even the slightest and the request for more will be even greater.
Matt Buchanan. Giz Explains: GPGPU Computing, and Why It'll Melt Your Face Off
URL: http://gizmodo.com/5252545/giz-explains-gpgpu-computing-and-why-itll-melt-your-face-off
---
As mentioned in the introduction, chapter 5 discusses power conservation on a small component scale, all the way up to conservation challenges of warehouse scale computers (WSCs). The first term analyzed is PUE; power usage effectiveness, which compared the IT power to total building power for a WSC. Several factors contribute to PUE, the majority of which would be the IT equipment and the cooling systems for the IT equipment (About 60-65% combined on average – Figure 5.2). One of the interesting things about the centers that have a good consumption ratio (under 2:1), interestingly enough, is because of non-computer related improvements such as air flow efficiency around the cooling systems and power conversion loss reductions. It was also interesting to note that with multiple components of any given data center, idleness is discouraged by spreading a small system load across multiple components (and vice-versa with large loads) rather than on one or a few, because it conserves more energy. This is true because a system will consume energy even on idle, but the increase of consumption with a light load is minor, making it a more efficient use to utilize multiple systems rather than a single system. The disk alone of a system can spend as much as 70% of it's energy alone keeping the hard drive platters spinning (Pg. 71). As I noted previously regarding technological improvements with GPUs, computing speed itself is nearly doubling at a rate of every 18 months, while maintaining the same power ratio. The article mentions that disk and DRAM are not on pace, however, and that future (well) balanced systems may become storage dominated, which will only increase the inefficiency of most of the world's already poor WSCs. The possible, some being quite easy to implement, solutions seem to be well within grasp, however, and it should be interesting to see if companies follow in the foot steps of others (such as Google) and the growing demand for energy conservation.
Chapter 5 from "The Datacenter as a Computer An Introduction to the design of Warehouse-Scale Machines". Synthesis Series on Computer Architecture, Morgan & Claypool Publishers, May 2009.
