Project Coordinator (EU) :Chalmers University of Technology
Country of the EU Coordinator :Sweden
Project participants :
EU: Chalmers University of Technology
Ahmed Ali-Eldin Hassan is a PhD. researcher, and is leading an effort into cloud efficiency, and energy-efficient edge computing. He has been working on the power-performance tradeoffs of computing systems including on debloating machine learning (ML) systems. The research interests aim to reduce the resource consumption of cloud and ML systems by limiting the power allocated to an application and removing unnecessary features and inefficiencies in cloud and ML systems using automated tools.
US: University of Massachusetts, Amherst:
Dr. Shenoy and his colleagues have been funded by the NSF to design a "Carbon First" approach to (i) make carbon a first-class systems design goal and (ii) to decarbonize cloud computing through carbon-aware software optimizations. These two projects address complementary issues and will come together to experiment with making cloud software applications energy-efficient and zero-carbon.
State of US partner :Massachusetts
Starting date :
Reducing Carbon footprint of Computing Systems
As noted earlier, the exponential growth of cloud platforms and the resulting growth in energy usage will become unsustainable due to the high carbon cost of cloud workloads. The community has realized the need to ensure new research to ensure sustainable growth of cloud platforms. For example, research has shown that training a single large machine learning model can result in as much carbon emissions as those emitted by a car for an entire year.
Our hypothesis is that there are significant opportunities for making modern cloud application (such as ML training and others) carbon- aware, and employing carbon-aware software design that incorporates such opportunities is necessary to making cloud computing sustainable. Our systems and experimental research will demonstrate these issues through its emphasis on greening cloud computing.
The proposed research and experiment are as follows. First, we will consider several applications such as ML training, distributed "big-data" data-processing, and web-microservices to quantify their power-performance trade-offs. By studying the power, we will also learn about their energy and carbon footprint. The carbon footprint of an application depends on the carbon footprint of the electricity/energy supply, which varies by region and by country.
Our experiment will use NSF CloudLab sites in the US and the edge lab from AI Sweden in Sweden, along with different public EU cloud providers. We will run the same workload at different sites to measure energy and carbon usage and show how carbon usage can vary by region and over time. After performing this baseline experiment, we will consider how to reduce the carbon usage of cloud applications, potentially making them zero carbon. This will be achieved through multiple techniques. We will leverage our work (at Chalmers) on software de-bloating of large applications such as ML training to remove bloat which leads to resource waste. Further, we will use workload elasticity techniques (being studied at UMass) to dynamically adapt the carbon consumption of the de-bloated applications based on the carbon intensity.
For example, the resource usage can be elastically increased when the proportion of renewable energy in the supplied electricity is high. A service such as Electricity Map which exposes the carbon cost of energy in various parts of US and EU will be used to obtain the carbon data (we already have a collaboration with the Electricity Map service). We will then experimentally run these green versions of our applications on CloudLab and SICS North data-centre and demonstrate the reduction in carbon usage by these applications. The research will yield design principles for making cloud platforms and applications sustainable and low carbon.
Impact 1: Enhanced EU – US cooperation in Next Generation Internet, including policy cooperation.
- The project hosted many US top researchers in ApPLIED along with many of their EU counterparts. In addition, three young EU researchers have been hosted by UMass.
- The project's focus is on carbon-neutral computing. We believe that our results can be used by policy makers to enhance the discussions on how the NGI, 6G, and other systems can be built to be carbon-neutral. Towards this end, the PI at UMass has gotten a large research project on reducing the carbon footprint in computing. The PI at Chalmers has been selected as a Swedish Young Research Leader with a grant of 1.5M Euros where part of the work will investigate how to decrease the power consumption of computing systems using edge platforms. Both PIs are actively engaged in public discussions with their local authorities and companies.
Impact 3: Developing interoperable solutions and joint demonstrators, contributions to standards.
Our end goal is to build a joint demonstrator, with much better integration between our testbeds. While our testbeds are not the only used platforms for our testing due to the nature of our projects, we already have established testbeds and run experiments together. Our accepted paper to the WebConf has used resources at UMass. Our work on de-bloating ML systems has used a recently setup testbed at Chalmers. Our intention for the project continuation is to be able to eventually run experiments across both sites for joint demonstrators.
Impact 4: An EU - US ecosystem of top researchers, hi-tech start-ups / SMEs and Internet-related communities collaborating on the evolution of the Internet
We hosted a panel in ApPLIED to discuss these topics, including how the evolution of the Internet, and of distributed systems in general will be over the coming decade.
The project's focus is on carbon-neutral computing. More results and how they were obtained can be found in the deliverable. We note that this work is still in progress. In particular, we are still working on running the measurements using the setup we have finally gotten running at Chalmers. This work will be done on local faculty funding.
We believe that our results can be used by policy makers to enhance the discussions on how the NGI, 6G, and other systems can be built to be carbon-neutral.
Future Plan :
Key results from the experiment include:
- Co-organized ApPLIED, the workshop on Advanced tools, programming languages, and PLatforms for Implementing and Evaluating algorithms for Distributed systems, colocated with PODC, a top conference in distributed systems.
- Invited paper at ApPLIED, (DARTS: Distributed IoT Architecture for Real-Time, Resilient and AI-Compressed Workflows)
- Accepted paper at the ACM WebConf, ?
- Invited to a dagstuhl seminar