Efficient Workload and Reference Management
A thesis submitted in conformity together with the requirements
intended for the degree of Doctor of Viewpoint
Graduate Department of Electrical and Laptop Engineering
College or university of Barcelone
c Copyright laws 2013 by simply Hong Xu
E cient Workload and Reference Management
Doctor of Idea
Graduate Section of Electrical and Computer Engineering
College or university of Toronto
This kind of dissertation concentrates on developing methods and systems to improve the e ciency of functioning mega datacenters with thousands and thousands of servers. In particular, it seeks to cope with two problems: First, tips on how to distribute the workload among the list of set of datacenters geographically used across the large area? Second, how to control the server resources of datacenters using virtualization technology? In the ﬁrst part, we all consider the workload management problem in geo-distributed datacenters. We ﬁrst present a novel distributed work load management protocol that collectively considers ask for mapping, which in turn determines how you can direct end user requests to an appropriate datacenter for digesting, and response routing, which will decides how to select a course among the pair of ISP links of a datacenter to path the response packets back in a user. Over the following chapter, all of us study some key facets of cost and workload in geodistributed datacenters that have certainly not been completely understood prior to. Through considerable empirical research of weather data and cooling systems, we make an instance for temp aware work load management, where the geographical variety of temperatures and its effect on cooling strength e ciency can be used to reduce the overall chilling energy. Additionally, we endorse for healthy workload management for both equally interactive and batch jobs, where the delay-tolerant elastic character of set jobs can be exploited to increase reduce the strength cost. A regular 15% to 20% cooling down energy reduction, and a 5% to ii
twenty percent overall cost reduction happen to be observed via extensive trace-driven simulations. Inside the second section of the thesis, we all consider the resource administration problem in virtualized datacenters. All of us design Core, a scalable and ﬂexible architecture that e ciently supports a number of resource managing policies. We implement a prototype of Anchor over a small-scale under one building datacenter with 20 machines. Experimental results and trace-driven simulations demonstrate that Anchor is e↵ective in realizing various useful resource management guidelines, and its straightforward algorithms are practical to fix virtual machine allocation with thousands of VMs and machines in just ten seconds.
to my parents
The ﬁrst the first is easy: Baochun Li. I actually started working with him as being a Master student back in 3 years ago. Throughout the years his eagerness and vision spur my personal curiosity toward many di↵erent topics of computer networking, great patience and revision work polish up nearly all single newspaper of my very own. His nontechnical advice as well serves well at various elements, such as opportunities and musical technology instruments. Of course there is also the regular workout for Hart Property that better my 2km jogging record many times. We couldn't overstate my friend Chen Feng's contribution to this work. I actually beneﬁted hugely from various discussions exactly where my difficult thoughts and ideas can be challenged and completed with his clear information and wide knowledge about mathematics. His e↵orts cause technical evidence of the concurrence of the m-block ADMM protocol. And actually ADMM was brought to me if he was posting Emmanuel Candes's talk in the group.
I actually am as well grateful for several members of BA4176 who also added more ﬂavors to my sixyear stay in this article. The list with the order of seniority: Mea Wang (with her Beagle Snoopy), Chuan Wu, Xinyu Zhang, Jin Jin (with his partner Amber and son Aaron), Hui Wang, Di Niu, Zimu Liu, Jiahua Wu, Elias Kehdi, Boyang Wang, Yuan Feng, Wei Wang,...
Bibliography: NSDI, 2010.
putting jobs in cooling-e cient locations inside the data middle, ” in Proc. USENIX
ATC, 3 years ago.
 G. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computation: Numerical Methods. Athena Scientiﬁc, 1997.
 S. Boyd and M. Vandenberghe, Convex Optimization. Cambridge University Press,
and services research, ” in Proc. USENIX HotCloud, 2009.
 Z. Cao, Z .. Wang, and E. Zegura, " Functionality of hashing-based schemes to get Internet load balancing, ” in Proc. IEEE INFOCOM, 2000.
Ej. Netw. Manag. (IM), 2009.
Manag. (IM), 2011.
 Y. Chen, A. Dieses, W. Qin, A. Sivasubramaniam, Q. Wang, and And. Gautam, " Managing machine energy and operational costs in hosting centers, ” in Proc. ACM Sigmetrics, 2005.
 J. Leader, " Beneath the covers in google: Current systems and future guidelines, ”
On the internet I/O, 08.
 M. Dean and S. Ghemawat, " MapReduce: Simpliﬁed data processing about large clusters, ” in Proc. USENIX OSDI, 2004.
Dropbox: Understanding personal cloud storage companies, ” in Proc. ACM IMC,
precision cooling system sales
brochure, ” http://tinyurl.com/c7e8qxz, 2012.