Topics: Data center architectures; Cloud reliability and performance
Authors: Esa Hyytiä (University of Iceland & Aalto University, Iceland); Rhonda Righter (University of California Berkeley, USA); Jorma Virtamo (Aalto University, Finland)
Presenter Bio: Dr. Esa Hyytiä received the M.Sc. (Tech.) degree in engineering physics
and Dr.Sc. (Tech.) degree in electrical engineering from Helsinki
University of Technology (TKK) in 1998 and 2004, respectively. During
2005-2006, he was working as a postdoc researcher at Norwegian
University of Science and Technology (NTNU), Norway, and 2006-2009 as a
senior researcher in Telecommunication Research Center Vienna (ftw.),
Austria. From 2009 until 2015, he worked as a senior research scientist
at Aalto University, Finland. Currently, he is Associate Professor in
the department of computer science at the University of Iceland. His
research interests include performance analysis, design and optimization
of various computer and communications systems.
Abstract: We consider single- and multi-server systems, where jobs have a maximum
waiting time (deadline) defined, e.g., by a service level agreement. A
fixed cost is a cost associated with deadline violations and the task is
to minimize the long-run cumulative costs. Job sizes (service
durations) are observed upon arrival, and current queue backlogs are
known. For a single FCFS server, the optimization task is to find the
optimal admission policy that may reject a job upon arrival if admitting
it would cause in future one or more deadlines to be violated (in
expectation). For parallel FCFS servers, the policy must (i) either
accept or reject a job upon arrival, and if accepted, (ii) assign it to
one of the servers. We derive efficient deadline-aware policies in the
MDP framework. For a single server, we obtain the optimal admission
policy. For dispatching to parallel servers, we develop efficient
heuristic admission and dispatching policies, whose performances are
evaluated by means of numerical examples. Additionally, we give some
exact closed-form results for heavy-traffic limits.
