Monthly bulletin of the IEEE Computer Society Special Technical Community on Sustainable Computing
Providing quick access to timely information on sustainable computing.
by Christopher Stewart, The Ohio State University
ball Association (NBA). Perhaps surprisingly, their coach, Bill Russell, was in only his second season as a coach in 1968. Even more shocking, coach Russell also played center. He is the only player-coach in NBA history to win a championship.
Well, I am now a second-year editor for the Sustainable Computing Register. I want to do more than edit. Starting in the next issue, I will contribute a regular column on economic and policy incentives for sustainable computing. While my business cards say distributed systems, economics and policy have always interested me. I doubt that I will be as prolific of a writer as Bill Russell was a center, but I hope you all will take a look at my columns from time to time. Sustainable systems, like no other technology-related field, are heavily influenced by national and global policies.
Also, I want to remind all members to get involved with Pick of the Month voting. Our process changed this year to increase participation. Anyone reading this letter can visit http://www.cse.ohio-state.edu/~cstewart/potm and vote. The process is set up to encourage self-interested behavior. Specifically,I want to highlight the following incentives: First, you should nominate your own papers. Second, you
should vote for your own papers (no rule against that). Third, you should cajole your colleagues to vote for your papers (stuff the ballot box). Fourth, you should cajole other voters to vote for your papers (explicitly encouraged via a comments section). Finally, if your paper wins, you should vote other really good papers (keeping the prestige of the award high). The Pick of the Month highlights great work in the field of sustainable computing because we publish in a diverse range of conferences. We need your participation and passion.
The Pick of the Month series highlights top research papers and projects within our community. It is com-
munity-driven award with an open nomination process and selection.
Anyone can nominate a paper by simply sending an email to either Prof. Christopher Stewart or Prof. David
Chiu. Then vote and converse about your nomination online by:
1. Visiting the URL below and clicking through to the Pick of the Month survey
2. Posting (anonymous) comments describing why the paper deserves pick of the month recognition. You
can even comment about why one nominee is better than another.
3. Every month, nominees with the most votes are eligible for Pick of the Month. SC STC Information Of-
ficers will decide among eligible candidates.
Try it out now. http://www.cse.ohio-state.edu/~cstewart/potm
We resume our Pick of the Month series with Greening Geographical Load Balancing, a SIGMETRICS 2011 paper that has been incredibly impactful. The paper builds on a now widely accepted trend: When datacenters scale out, the spread out across geographically diverse locations. Amazon, Google, and Microsoft have datacenters in almost every corner of the globe. Even large enterprise companies in banking and insurance are operating datacenters multiple locations.
Geographically distributed locations offer the benefits of diversity. One location is often unlike another. At the time that our pick of the month was published, datacenter managers knew that they could exploit different geographic and temporal pricing differences. The Pick of the Month showed that 1) naively routing work for low cost could increase energy and carbon footprints and 2) wisely routing work for low cost and low energy keeps costs low in a sustainable fashion. Note, the word "showed" understates the paper's rigor. In this issue, we sit down with the lead authors, Zhenhua Liu and Minghong Lin.
Christopher Stewart: This work is becoming one of the most cited works in the field of sustainable computing. When I looked, Google Scholar estimated over 70 citations in just 2 years. What are your thoughts about the community's response to this work? Why has it been received so well?
Zhenhua and Minghong: We are really happy to see that the community appreciates our efforts. The idea of using geographical load balancing to extract benefits from heterogeneity is widely used in subsequent papers, e.g., to reduce power cost and/or GHG emissions, to ease the incorporation of renewable energy. Some following work also extended our model, which is very exciting. The reason why it has been received so well is, in our mind, that we tried our best to provide a solid model together with practical distributed algorithms. Together with the GreenMetrics paper “Geographical load balancing with renewables” (best student paper), we provided numerical results to quantify the benefits and identify the potential of geographical load balancing in demand response, as verified by subsequent works.
Christopher Stewart: OK. We'll publish the plug to your GreenMetrics paper. (laughing) Mainly because it is good work as well and is also a pick of the month nominee. Let me ask about the process of innovation for you guys. As the title suggests, this paper builds on prior work on geographical load balancing by adding a “green” element. The key observation is that load balancing for low energy costs actually wastes energy. To me, this is a really clever observation. How did you uncover this problem?
Zhenhua and Minghong: This was first uncovered from our model. By constructing a reasonable model and characterizing the problem, we found this seemingly surprising result. Interestingly, we noticed that similar phenomenon has been revealed in other area, e.g., Jensen’s paradox, which states the demand, is expected to increase if the cost decreases. We then conducted numerical simulations to validate our observation. Therefore, we think mathematical models are very useful to provide insight into system design.
Christopher Stewart: Students in my special topics courses often miss the importance of distributed solutions. Can you explain the virtues of distributed load balancing compared to a centralized approach?
Zhenhua and Minghong: Generally, distributed load balancing algorithms can provide better robustness, faster response, lighter communication and synchronization overhead, and other benefits. The benefits are more significant for a system with a large number of points, e.g., Akamai. Even for Google, from our collaborations with them, distributed solutions are usually preferred if possible and they actually make a huge effort on this and sometimes they are pushed by the system scale to transit to distributed solution. From the theory side, our distributed algorithm extended existing gradient-based method and guaranteed convergence even when the objective function is not smooth enough.
Christopher Stewart: Your paper includes an entire subsection labeled “The Importance of Dynamic Pricing.” Your paper helped my understand the importance of dynamic pricing in sustainable computing systems. Can you talk a little about this mechanism and why it is useful?
Zhenhua and Minghong: Before we have some emission regulation or carbon tax, it is really hard, if not impossible, for data center operators to focus on reducing emissions other than electricity bill, although some big companies such like Apple and Google did this due to increasing public pressure and future energy security consideration. Among different options, e.g., coincident peak charging, or ancillary market participation, one possible approach is real-time dynamic pricing. In our paper, we illustrated by numerical simulations that the
non-renewable energy usage could be significantly reduced if real-time electricity pricing was used proportionally to the instantaneous non-renewable energy generation. We also pointed out if the pricing was not set in this way, the reduction can be limited, which was later supported by the Sigcomm 2012 paper “It’s not easy being green”. Therefore, we really hope the transition to smart grid can provide more incentives for customers such as data centers to utilize their demand flexibilities to act as “virtual storages” through various demand response programs.
Christopher Stewart: Thank you both for the excellent paper and for this insightful interview. I fully expect to interview you again since this paper is part of bigger body of excellent work. Congratulations.
This month's feature is an overview of the PREFiguRE (Performance, Power, and Reliability
Framework for Disk Drives) framework for managing disk reliability and power.
Contributing authors from left to right: Feng Yan, Xenia Mountrouidou, Alma Riska, and Evgenia Smirni
PREFiguRE: As data centers grow bigger and central to many service providers and enterprises, storage systems have drawn attention. Despite the emerging new storage technologies such as solid state drives (SSDs), hard disk drives (HDDs) still host the overwhelming majority of enterprise and consumer data. Specifically, hard disk drives continue to store most of the aging data, which are expected to grow in size over the years. While data center power consumption amounts to about 20% of all IT power consumption, as much as 20% to 40% of data center power consumption is attributed to disk drives. Because not all data in a vast data center is accessed simultaneously, a compelling approach for reducing power consumption in data centers is to spin down idle hard disk drives. This approach is routinely deployed in storage systems that serve as archival or backup systems and is being exploited even in high-end computing environments.
Spinning down disk drives to save energy is a challenging open problem because it can impact user performance and disk reliability. If requests arrive while the disk is still in a power saving mode, they can be delayed by the disk reactivation process. In addition, even though the disk drives are under utilized, the idle times can be highly fragmented. Idle periods that are long enough for efficient power savings are limited. Finally, every power up/down wears out the disk drive. As a result, disk drives should be put into powersaving modes only for a limited number of times to prevent premature wore out.
Common practice methods try to address these challenges by idle waiting for a fixed amount of time or guide scheduling decisions based on the past utilization. However, none of these methods provides performance guarantees or takes into consideration disk reliability. PREFiguRE overcomes these short-comings. In its core, there is a framework that uses as input user- or system-level constraints such as the number of allowable power ups/downs of a disk within a time period (strict constraint) and the user acceptable potential performance degradation of future IOs (soft constraint), and estimate the projected power savings as well as provide a strategy on when and for how long these power savings should be scheduled. The histogram of past idle times plays a key role in PREFiguRE as it projects future power savings based on statistical information that is monitored or extracted from this histogram. Probabilistic interpretation of all of the above information leads PREFiguRE to define robust schedules for power saving modes. PREFiguRE captures the workload changes in the system by updating the histogram of idle times as well as information about the sequence of idle times. Such updates enable the adjustment of the schedules of power saving activation to the workload dynamics.
We have evaluated PREFiguRE on enterprise disk drive traces with a wide range of idleness characteristics. We have seen excellent agreement between the results from PREFiguRE's analytic estimations and the trace driven simulations, suggesting it has practical benefit.
Officers reports from April 1, 2013 to April 29, 2013
Report from Secretary/Treasurer (Danilo Ardagna):
• Collected officers' activity reports and prepared monthly STC report
Report from Conferences Chair (Diwakar Krishnamurthy):
• Solicited collaborations from organizers of DCPERF, USENIX ATC, HPDC, ICS, SIGMETRICS, ISCA, WETICE, IGCC, GCN, IWQoS
• Continuing to update list of conferences related to STC
Report from Academic Chair (Niklas Carlsson):
• Identified and contacted candidates for future community highlight features
Report from Membership Chair and vice-Chair (Sergey Blagodurov, Matthew Forshaw):
• Counted the number of members every week for the past month.
• Sending invitations to potential STC-SC members and continuing working on the new invitation list
Report from Communications Chair (Abhishek Chandra):
• Prepared the upcoming events for announcement with the communications vice chair
Report from Policies and Procedures Chair (Stephen Dawson):
• Contacting colleagues to find a vice-chair
Report from Industry Chair (Canturk Isci):
• Working with the academic chair to identify the next set of community highlights
• Starting work on a next industry highlights feature
Report from Information Officers (Danilo Ardagna, David Carrera, Fan Dongrui, Guillaume Jourjon):
• Contributed material for newsletter and blogs
Report from the Newsletter Editor (Christopher Stewart):
• Produced April newsletter
• Set up public website for Pick of the Month
Report from the Web Master (Yan Shvartzshnaider):
• Added a resource page including links to relevant websites and interesting articles
By Abhishek Chandra, University of Minnesota and Bhuvan Urgaonkar, Penn State University
Chair:Anirban Hahanti, NICTA
Policies and Procedures: Stephen Dawson, SAP
Secretary/Treasurer: Danilo Ardagna, Politecnico di Milano
ndustry Chair: Canturk Isci, IBM
Conferences: Diwakar Krishnamurthy, University of Calgary
Editor: Christopher Stewart, Ohio State University
Conferences: Amarjeet Singh, IIIT-Delhi
Editor: David Chiu, Washington State University Vancouver
Academics: Niklas Carlsson, Linköping University
Information: David Carrera, UPC BarcelonaTech
Membership: Sergey Blagodurov, Simon Fraser University
Information: Danilo Ardagna, Politecnico di Milano
Membership: Matthew Forshaw, Newcastle University
Information: Fan Dongrui, Inst. of Computing Technology
Communication:Abhishek Chandra, University of Minnesota
Information: Guillaume Jourjon, NICTA
Communication: Bhuvan Urgaonkar, Penn State
Web Master: Yan Shvartzshnaider, University of Sydney