Course Info

Energy is an important resource for today's computer systems ranging from laptops (think battery life) all the way up to entire data centers and clouds (staggering electricity bills). There has been a lot of recent research on the topic of improving energy-efficiency. This research cuts across different areas of computer science including Computer Architecture (ex: energy-efficient processor design), Networking (ex: putting the network to sleep), Storage (ex: exploiting idle I/O periods to spin down disks), Distributed Systems (ex: capacity provisioning), and Modeling/Analytics (ex: modeling a server's power consumption). This course will cover the recent research trends in energy-efficient computing, focusing specifically on themes and techniques for achieving energy-efficiency in computer systems.

This is a seminar course and will involve research paper discussions and a semester-long project.
Students should read the assigned papers (on average, 2/class) and submit a hard copy of their review for the non-optional paper. The reviews are due before the class begins.
Most, if not all, of the paper discussions will be lead by the instructor. However, students are expected to voice their opinions on the papers to make the class more interactive (=fun).

Grading will be based on submitted reviews, class interaction, and the project.

Syllabus & Schedule

Date	Topic	Readings	Notes
Jan 26 (Tuesday)	Introduction, Class logistics	None (yay!)	We will discuss class logistics and overview of syllabus.
Jan 28 (Thursday)	Energy-efficiency basics	energy_proportional, EPA Review not required :)	Scan the EPA report for facts and results, skip the non-technical stuff.
Feb 02 (Tuesday)	SPEED: processor	reduced_CPU (optional), opt_allocation Read the reduced_CPU paper first.	For the opt_allocation paper, ignore the appendices and skip over the proofs. However, read the theorem statements carefully and try to understand what they are saying about the optimal allocation.
Feb 04 (Thursday)	SLEEP: full-system	power_nap, agile (optional) Read the power_nap paper before the agile paper.	For the agile paper, don't worry much about sections 4 and 5, and just scan through them quickly.
Feb 09 (Tuesday)	SCALING: stateless vs stateful	autoscale, cachescale (optional) Read the autoscale paper first.	For the autoscale paper, Section 4.3 is quite technical. Feel free to scan through it quickly.
Feb 11 (Thursday)	No class		Career fair day.
Feb 16 (Tuesday)	SHARING: distributed vs local	energy_routing (optional), softscale
Feb 18 (Thursday)	SMARTS: custom-made systems	fawn, heteromates (optional)
Feb 23 (Tuesday)	No class		Anshul is away at a workshop in DC.
Feb 25 (Thursday)	SOURCE: renewables	sustainable_energy (optional), renewable_cooling
March 01 (Tuesday)	SLEEP: processor	dreamweaver, barely_alive (optional)
March 03 (Thursday)	Project idea presentation		Each group will give a short (5-8 mins) presentation about their project idea.
March 08 (Tuesday)	SLEEP: network	idle_net (optional), somniloquy sleepless (optional) Read the idle_net paper first.	The sleepless paper is optional. It is closer to the idle_net paper but the evaluation is done on a larger, more realistic system.
March 10 (Thursday)	SLEEP: storage	maid, pdc (optional)
March 22 (Tuesday)	SLEEP: memory	power_page, commercial (optional)	Mid-semester grades out.
March 24 (Thursday)	SLEEP: sensors	sensor_power, solar_embedded (optional), sensor_detect (optional)
March 29 (Tuesday)	SCALING: setup time	napsac, vmsetuptime (optional)
March 31 (Thursday)	No class: Finalize project setup.	proposed meeting schedule	Meet with instructor to discuss and finalize project setup.
April 05 (Tuesday)	No class		Anshul is away at a workshop in Germany.
April 07 (Thursday)	No class		Anshul is away at a workshop in Germany.
April 12 (Tuesday)	SCALING: storage	rabbit, sierra (optional)
April 14 (Thursday)	SCALING: memcache	mem_hotspots, dynacache (optional)
April 19 (Tuesday)	SCALING: survey	autoscaling_survey (optional) (review not required, but scan through it)	This is a (fairly extensive) survey paper on autoscaling strategies.
April 21 (Thursday)	SPEED: power	power_capping, power_modeling (optional)	Feel free to quickly scan through the control theory part (Sections 4 and 5). power_modeling deals with how to model server power consumption and, as such, does not really fall under the "SPEED" category.
April 26 (Tuesday)	SPEED: memory	memscale, memory_dvfs (optional)	These are very similar papers, and they were published around the same time.
April 28 (Thursday)	SPEED: disks	disk_energy, drpm (optional)	disk_energy looks at four techniques, while drpm explores one of them in depth.
May 02 (Tuesday)	review	power_provisioning, power_struggles (optional)

Resources

• How to read a research paper, How to write a technical paper review, Reading a CS research paper
• Project ideas (talk to Instructor to figure out potential resources available for each project):
  • SPEED:
    • What is the optimal DVFS setting for a given workload?
    • How does optimal DVFS setting change with workloads?
    • What is the optimal DVFS setting for a given workload on a mobile device?
  • SLEEP:
    • C-state power management on servers.
  • SCALING:
    • Autoscaling for stateless tiers.
    • Autoscaling for Memcache.
    • Autoscaling for Database.
    • Comparison of autoscaling policies.
    • What should autoscaling react to (response time, queue length, request rate, etc.)?
    • Autoscaling in the presence of multiple applications.
  • SHARING:
    • How can we share database with other applications?
    • How can we share among replicas of databases?
    • How can we share among any two applications?
  • Figure out how system power consumption varies with resource utilization of individual system components such as CPU, Memory, N/W, etc. This will require exploring some monitoring capabilities on the target system.
  • Explore the optimal DVFS configuration for a distributed system. This will require setting up a distributed application.
  • Play around with various autoscaling policies, including those offered by Amazon.
  • Explore how stateful nodes (caching or storage) can be scaled. This is a difficult problem but can easily result in a research paper.
  • Explore how to optimally share load between heterogeneous servers.
  • Explore how non-grid energy sources (solar, wind, etc) can be integrated with the grid to handle arrival rate variations in workloads. This will largely involve simulations or numerical experiments.
  • Which workloads benefit more from DVFS? Investigate via experiments and by exploring various workloads.
  • Explore the various power saving features on a typical laptop or server, and figure out how to pick the right option based on various workloads or load conditions.
  • Any other project along the above lines.

Grading (tentative)

• Paper reviews: 30%
• 1-page review for each assigned paper (hard copy only, please). No team work allowed on this.
• The review is basically what you think the paper was about and what you learnt from it, in your own words, especially in the context of energy-efficiency. The review will be graded on a 10-point scale, and should contain the following five sections:
  • Summary (4/10 points). 2 paragraphs (8-10 sentences) summary of the paper.
  • Strengths (2/10 points). At least 2 significant strengths of the paper.
  • Weaknesses (2/10 points). At least 2 significant weaknesses of the paper.
  • Technique (1/10 point). Mention the energy-efficiency technique(s) employed by the paper. Your choices are:
    • Speed (uses device-dependent active states to slow down the device and reduce power)
    • Sleep (uses device-dependent inactive states to conserve power)
    • Scaling (turning devices on or off dynamically in a multi-device setting)
    • Sharing (consolidating multiple different workloads on to a single device)
    • Smarts (using customized devices)
    • Source (using renewable energy sources to offset grid energy or reducing thermal effects)
    • Other (if different from aboave, explain)
  • Metrics (1/10 point). What are the main metric(s) employed in the paper, for both power/energy/cost and performance.
• Reviews are due at the BEGINNING of each class.
• Reviews should be original work and should be done individually by each student.
• You are allowed to drop one lowest paper review score in the semester.
• You can submit paper reviews in advance if you know you will be missing class in the future.


• Class interaction: 15%
• The basic idea is to get you to attend classes, participate in paper discussions in class, and actively participate in the project discussions.
• By the end of the semester, if I can recognize you based on your contributions to the class and project discussions, you should get a good score on this.


• Class project: 55%
• The project will be graded out of 100 points as follows:
  • Project proposal presentation (5 points)
  • Final project presentation (15 points)
  • Project report (15 points)
  • Project progress (discretionary, depends on project deliverables and range of evaluation) (65 points)
• This should basically be a (systemsy) research project, done in groups of 2-4. Some project ideas are listed in the Resources section. If you have an idea of your own that deals with "energy-efficiency", broadly defined, that is perfectly fine (and encouraged even).
• By the end of 3-4 weeks, you should have decided on a project and a project group. We will have one lecture slot devoted to short, 5-minute presentations on each project. There will also be an update presentation by each group around the middle of the semester. And then a final presentation towards the end of the semester.
• Project report: You will also have to submit a 5-10 page research-paper quality project paper (one per group) at the end of the semester (Dec 16th deadline). You can use Word, Latex, etc. (no contraints). I would suggest having at least the following sections in the report:
  • Introduction: Start with some motivation (maybe cite some papers or numbers, if it helps), state your problem statement and goal, and maybe briefly summarize your results. At the end, mention how the rest of the paper is outlined. See the autoscale or opt_allocation papers for a good Introduction layout.
  • Experimental setup: Mention what machines/instances you are using, whether you are using VMs or physical machines, or both. Mention what application/benchmarks you are using, and what metrics you are using, and how you are measuring them. These can all be in different subsections.
  • Approach: Discuss your proposed approach, and why you think it is a good idea. If you have multiple approaches/algorithms you will be using, discuss all of them. You can also expand on your problem statement here if you wish. Also mention which strategy you are using (SPEED, SLEEP, SCALING, etc).
  • Results: This is where you put most of your evaluation results. This section should be graph-heavy. The more the merrier. Explain the plots in text, and also have meanigful captions. Also, for graphs, please make sure to have legible legends, axes titles, and axes values. A common mistake is to have unreadable (very small sized) text in graphs.
  • Prior work: Have sufficient (at least ten) references in this section. These should be papers that are addressing the same problem as you are. It is ok if they are doing similar work. Make sure you mention the commonalities and differences between your paper and the cited works.
  • Conclusion: Summarize your problem, approach, and prior work. Then summarize your main results. Also mention possible future work.


• DO NOT CHEAT! I will make a reasonable effort to make sure that you are writing your paper reviews honestly (by yourself) and that you are working on the class project without borrowing all the ideas from existing literature. If you need any further clarifications, please talk to the instructor.