Brief summary of Session 4 of HotNets VII (15:30 Monday, October 6, 2008).

Title of session: Cloud and Data Centers
Chair: Martin Arlitt (HP Labs, U of C)

Paper 9:  Plugging Into Energy Market Diversity, Asfandyar Qureshi (MIT)
Speaker:  Asfandyar Qureshi

This talk focused on how to reducing the energy cost at large distributed 
data centers.  Based on the observations that (i) energy consumption vary 
per day and location, (ii) large distributed systems typically have 
replicas located at different geographic locations, and (iii) computation 
can be moved to different locations, the authors consider a replica 
selection problem aimed at minimizing the energy cost needed to serve a 
fixed load.  (The model does not take into account peak vs. average 
demands, and assume one-day-ahead-knowledge, of energy prices.)  Using 
statistics of the energy cost in a number of different US cities, the 
author showed that variations in energy cost allow dynamic replica 
selection techniques to significantly outperform static replica 
allocations.

During the question and answer session of this talk a number of 
interesting issues were raised.  It was suggested that the current model 
could be extended to take into consideration the cost of transferring data 
between data centers.  In addition, there may be additional issues related 
to many companies switching back and fourth between locations.  For 
example, it was suggested that if lots of companies used these techniques, 
the choices made by these companies might directly affect the cost of 
energy and therefore cause further oscillations in energy prices.

Paper 10:  On Delivering Embarrassingly Distributed Cloud Services, 
Kenneth Church (Microsoft), Albert Greenberg (Microsoft), James Hamilton 
(Microsoft)
Speaker: Kenneth Church

This talk discussed using large containers with server equipment to 
deliver embarrassingly distributed cloud services.  While it was noted 
that large data centers sometimes are beneficial, the author suggested 
that it often is beneficial to use multiple (large) data containers (at a 
cost of $2M, for example) rather than a few (hug) data centers (at a cost 
of $1B, for example).  Distinguishing between "mega" solutions (e.g., 
large data centers that must be custom built) and "micro" solutions (that 
uses some set of containers that can be bought pre-built, for example), 
the speaker provided insights with regards to when micro solutions are 
beneficial.  As containers relatively easily could be distributed at 
different geographic locations, the author suggested that micro solutions 
typically have a significant advantage in terms of geographic locality (as 
exemplified using a distributed spam filter application) and redundancy, 
and can be deployed much quicker (to react to fluctuations in demand, for 
example).

During the question and answer sessions the author further suggested that 
(i) containers typically would be built for a particular customer in mind 
and organizations typically prefer not to share containers/resources, (ii) 
organizations only should use "mega" solutions if the demands etc. are to 
large to be satisfied with some number of "micro" solutions, and (iii) the 
containers have enough redundancy built in that they do not need any 
physical human maintenance during their deployment period (of three years, 
for example).  The basic idea is that these containers can be manufactured 
at a low cost, and then shipped and deployed anywhere at a relative short 
notice.

Paper 11: Dr. Multicast: Rx for Datacenter Communication Scalability, Ymir 
Vigfusson (Cornell University), Hussam Abu-Libdeh (Cornell University), 
Mahesh Balakrishnan (Cornell University), Ken Birman (Cornell University), 
Yoav Tock (IBM Haifa Research Lab)
Speaker: Ymir Vigfusson

After having noted that most data centers do not use IP-multicast, and 
provided some examples for why (e.g., routers/switches, broadcast storms, 
and bad reputation in general), the speaker described their system called 
"Dr. Multicast" which maps traditional IP multicast operations to either 
use a new point-to-point UDP multi-send operation or to a traditional 
multicast address.  Further, their protocol is designed to respect any 
administrator-specified (acceptable-use) policies, and optimize resource 
allocations.  In general, the talk follows the general outline of the 
paper and discusses their general design (using a library module, mapping 
module, gossip layer, and optimization of resources usage using a 
black-box model), and the suggested benefits (central policy control, 
transparency, performance, and robustness).

During the question and answer period it was noted that a star topology 
was used for the experiments, and the author did not believe there was any 
other risks to the protocol than an administrator having to switch back to 
unicast (slowing down the applications to the current speed).