1. Green Computing Conclusions Tarek Abdelzaher Dept. of Computer Science University of Illinois at Urbana Champaign

2. What We Tried to Achieve A combination of depth and breadth: In-depth look at data center energy consumption Broad look at related energy problems A combination of instructor-driven and student-driven approach Depth coverage driven by instructor Breadth coverage driven by students

3. What You Should Expect Now Should be well-familiar with the state of the art in energy management of large computing systems Should be able to write a good paper on the topic Should be able to tell a novel problem from a solved problem Should be aware of general trends in energy research in other areas (transportation, smart buildings, green Internet, smart grid)

4. Examples of Publishable Future Work Project 1: Energy-efficient Map/Reduce Novelty claim: First paper to consider DVFS as a means to meet temperature constraints in the context of map-reduce energy optimization. Target community: systems, energy, autonomic computing Other approaches to energy savings relied on consolidation and turning machines off (or running machines at full speed then turning them off). Raising temperature such that full speed execution violates temperature constraints (fixed by DVFS) saves on cooling

5. Examples of Publishable Future Work Project 2: Green HDFS Novelty claim: First paper to consider adaptive partitioning of server farms into hot and cold servers (those hosting hot and cold data sets) to save energy. Target community: systems, energy, autonomic computing Other approaches that partition map-reduce application data into hot and cold used static data-to-server assignment (hence static partitioning into hot and cold servers). Dynamic partitioning servers can be more effective in dealing with load fluctuations.

6. Examples of Publishable Future Work Project 3: Energy-proportional Key-value Store Novelty claim: First paper to implement an energy-proportional peer-to- peer storage/DHT system Target community: P2P, energy, middleware Previous P2P solutions for distributed storage did not address energy-proportionality Energy-proportionality reduces the total consumption of P2P systems when not operating at peak load

7. Examples of Publishable Future Work Project 4: Green Memory Novelty claim: First paper to consider the pros and cons of unreliable (but more energy efficient) memory Target community: systems, energy, architecture With the right “killer app” (e.g., caching), if proven to be more resource-efficient, green memory can revolutionize the industry. The premise is that energy spent to provide reliability on top of green memory is outweighed by its energy savings

8. Examples of Publishable Future Work Project 5: Combining Multi-DVS with DPM Novelty claim: First paper to combine dynamic power management (DPM) with MultiDVS capabilities in real-time embedded systems Target community: real-time systems There is no prior work on platforms with MultiDVS capabilities, with the exception of one paper. That single paper did not consider the possibility to turn machines off In workloads with high imbalance in resource demand, opportunities for turning machines off are limited by the bottleneck resource, leaving other resources (e.g., memory) underutilized and subject to additional savings with DVS.

9. Examples of Publishable Future Work Project 6: DVS for Synchronized WSNs Novelty claim: First paper to consider the adverse effects of DVS on clock synchronization in sensor networks Target community: sensor networks No prior work uncovered (much less explored) the effect that DVS has on clock de-synchronization To keep clocks synchronized while bounding clock sync overhead, one must appropriately bound DVS transitions.

10. Examples of Publishable Future Work Project 7: Low-power Design for Manycore Synchronization Primitives Novelty claim: Paper reduces energy overhead of synchronization primitives in high-performance computing Target community: high-performance computing (?), architecture (?) If one can show that synchronization energy in manycore machines is a significant fraction of the total, and that the cost of the new solution does not outweigh the energy saving benefits, the contribution makes sense.

11. Research Directions: Measurement Estimating energy consumption using software measurements, instead of monitoring hardware Much prior research, but current models remain inaccurate Accurate models of single-machine energy consumption

12. Research Directions: Energy Proportionality Building energy-proportional systems Routers Stateful file systems Easy to do for stateless, highly replicated systems, but not so easy for stateful systems with unique, heterogeneous resources

13. Research Directions: Energy Optimization and Control Mature area for single servers Also mature area on the computing side (reducing energy of computation) Not much done on joint energy optimization of computing and cooling (with thermal constraints), with the exception of a few papers covered in this class

14. Research Directions: Instability Challenges Instability in computing systems is not a well-researched area Opportunities lie in identifying interesting new instances of instability (self-reinforcing “vicious cycles”) in practical systems, and building clever automated solutions to avoid the problem

15. Research Directions: Green Programming APIs An emerging direction is to embody energy- saving mechanisms into language constructs and abstractions used by distributed systems Green map-reduce Green HDFS Green cache Green distributed file systems

16. Research Directions: Green Internet Very few papers investigate high-performance router architecture for green computing Current core Internet routers themselves look like “server farms” Current routers are not energy-proportional

17. Research Directions: Green Transportation Understanding the “energy slack” in transportation systems Savings by better navigation Savings by better driving habits Savings by altering the time of departure (within acceptable limits) Driver notification systems that exploit the above understanding for reducing energy

18. Research Directions: Green Buildings Understanding the “energy slack” in residential systems Smart subsystems: smart thermostats, smart lighting, … Smart appliances (supply-following load): refrigerators, washing machines, driers Smart energy storage (hybrid vehicles, cooling, etc)

19. Research Directions: Smart Grid Two main challenges: Move from a fixed distribution architecture to a market will millions of consumers who may also serve as suppliers (e.g., households with solar panels) Accommodate renewable energy sources (and hence unpredictability in supply)

20. Major Industrial Research and Development Initiatives Job opportunities Internship opportunities Who to contact? Just look at the industry authors on the reading list for a quick picture (feel free to contact me for other pointers or for letters of recommendation)

21. Remaining Reminders Final: Out today Due (by email) by Sunday night, Dec 12 th 30 simple multiple-choice questions that test high-level knowledge of reading list

22. Projects Reports are due by 11:59pm on Friday, Dec 17 th. Report should not exceed 10 two-column pages, in 10pt font. Report must present motivation, relevant background, novelty claims, related work (explaining the value added by your project), design, and evaluation. “Short and to the point” is preferred to “long and content-free”.

23. Next Steps and Deadlines I strongly recommend publishing a paper from your project Upcoming deadlines: Energy and data centers: Jan 8 th ICAC; March 18 th IGCC 2011 Sensor networks: Feb 4 th DCoSS 2011; April 8 th Sensys 2011 Real-time: Jan 23 rd ECRTS 2011

24. Survey Thank you for answering the poll on how to improve this class in the future!