The concept of datacenters powered by renewable energy sources, including wind and solar, is nothing new. Companies like Apple and others have built solar farms and worked out deals with the local grids to exchange energy back and forth, while others rely on batteries to capture and store renewable energy.
While all of this is noteworthy at scale, for smaller datacenter operators, it is one thing to install the equipment and another entirely to tune the workloads that run on these green machines. According to Thu Nguyen, whose recent career has been built around finding ways for systems to take full advantage of renewable energy sources inside the datacenter, the key to operating efficiently with green power comes down to recognizing how to schedule and plan workloads according to available power—all the while minimizing loss of stored reserves.
Nguyen and a small team at Rutgers University have been testing the idea that efficient use of resources balanced against renewable energy supplies can dramatically increase utilization and pull down the PUE rating of datacenters, particularly those with high performance computing applications. This is not as simple as outfitting a small datacenter with solar panels and conversion equipment, it means making sure the applications themselves, not to mention the middleware glue that holds the system together, are primed for variable energy supplies.
The result of this work is called GreenPar, which the team has released in early open source version before pushing out a new set of features built into bth SLURM and Hadoop/MapReduce, that can dynamically schedule jobs based on available energy supplies. As Nguyen told The Next Platform, while these power storage and retrieval concerns are of interest to the datacenter operators, for end users, the applications are what matters—no matter how they’re being powered. This means it is critical to have a balanced scheduler that takes into account workload and energy efficiency demands in tandem. He explains that even at smaller datacenters in research settings, “There are mixed job types, which fits with how the flow of energy happens. This means that for more demanding jobs, we want to maximize their performance during times when we have an excess of solar energy, for instance, and be able to remain efficient when renewable supplies are lower while still providing a certain level of response and completion times.”
Using 16 small solar panels, which supply between 3 and 3.2 kilowatts, the team has proven the ability to support their few racks of compute running at peak capacity, although these are not HPC systems with fast processors and interconnects. Even still, the dynamic scheduling around available resources and the needs of compute-intensive jobs, some of which can get better speedup than others, is the key to efficiency. Nguyen demonstrated that GreenPar led to better consumption of the renewable energy sources, a subsequent decrease in the “brown” energy (that coming from fossil sources) and most important, a reduction in the average runtime for the high performance computing applications running on the Atom-based machines. While this was a low-power datacenter to begin with, the automated layer could also foretell future job requirements as well as what the expected renewable load might be. Nguyen says “the results also showed that an online policy using information about job speedups, runtimes, and predictions of solar energy production can come close to matching an offline policy that additionally has perfect information about future job arrivals and green energy production.”
The real problem with taking advantage of renewable energy sources inside the datacenter is the variability of supply, which has an impact both on how jobs are scheduled, but at the macro level, what to do with the reserves once they are collected. As one might imagine, it’s difficult to count on the stability of power from renewable sources like wind and solar, so a secondary goal is to make sure it can be stored for efficient reuse. There are some hurdles here, however, when it comes to retrieving stored energy. For example, while it’s quite possible to store renewable-derived power in batteries, there is quite a bit of energy “lost in translation” and further, over time, these very expensive batteries become exhausted after long stretches of charges and discharges.
On the other side, for those who want to send the power back to the grid for later use, there are losses here as well—both in terms of the transmission and receipt as well as in costs. The backend infrastructure for smoothly handling grid back-and-forth at most power companies is woefully inadequate, with many grid operators ill-equipped to handle multiple requests for storage and retrieval—something that Nguyen says his team is focusing on next.
One of Nguyen’s colleagues from the GreenPar project is working with Microsoft on their datacenter operations, but the future of how this code will adopted into larger scale datacenters where it could make a big difference at scale is unknown. “When you think about it, there is really no incentive for a lot of these big datacenter operators to reduce their energy consumption in smaller ways like this beyond marketing purposes. They want to maximize their datacenter investments and increase revenue—that’s really what drives them.” He notes that the national labs and some other smaller research centers are considering how renewable energy sources can fit into overall scheduling models, “but whether this could find a fit in industry is up in the air.”
One might expect that the big datacenter operators that have adopted some form of renewable energy use would have made it a priority to go beyond the solar panels and replacement of chillers with outside air, but at scale, this can be difficult to accommodate. Even though the hyperscale datacenters can have over 100,000 machines, these are basically “sub-clusters” and can be managed in a way that can take advantage of new middleware approaches that maximize utilization of compute and prime energy loads. This might make using a scheduling approach like GreenPar, which Nguyen says can likely scale to many thousands of machines