Traffic Control in Data Center Networks

PhD Thesis Proposal Defence


Title: "Traffic Control in Data Center Networks"

by

Ahmed MOHAMED ABDELMONIEM SAYED


Abstract:

Cloud computing defines a new computational approach that has led to the 
restructuring of the whole IT industry, by making computing available to 
customers at a reasonable cost in a man- ner similar to governments or 
private companies providing utilities (water, electricity, gas) to the 
citizens.

This new computing paradigm is becoming widely popular and increasingly 
replacing the legacy infrastructures within the public and private IT 
sectors. As a natural consequence, data centers have seen a dramatic 
increase in their rate of deployment energized by a sense of compe- tition 
among big Cloud Service Providers (CSP) such as Amazon, EMC, Google, 
Microsoft and Softlayer. It is widely believed that in the near future any 
piece of data or information that circu- lates the Internet originates in 
fact from some data center. Data centers consist of tens of thousands of 
servers mutually interconnected via high speed network interconnects, 
running a large number of applications that serve a huge number of users 
simultaneously. As such these applications typi- cally adopt a 
multi-tiered design model where several services residing on distributed 
servers work together to satisfy a single client request. Hence, the 
overall performance of such applications depends greatly on the ability of 
the underlying communication network to provide efficient and timely data 
transfers.

In this thesis proposal, first, we give the background of the 
architectural design of data center networks and inspect how such networks 
would affect application performance with more attention to network 
congestion and related work. Then, we explore and analyze the causes of 
performance degradation in such high-throughput low-latency Data Center 
Networks (DCNs) to provide guide- line in the design of an efficient 
congestion control scheme.

Motivated by the popularity of TCP as the major transport protocol in data 
centers, we pro- pose two novel TCP-AQM schemes to improve TCP efficiency 
without incurring any changes to the TCP protocol stack. First, we present 
an equal-share allocation AQM “RWNDQ” for DCN switches. RWNDQ relies on 
TCP flow control to efficiently and explicitly control TCP’s sending rate 
and hence achieve improved performance. Then, to address the popularity, 
scale and severity of Incast congestion in datacenters, we propose an 
Incast-Aware Queue Management “IQM” system which predicts possible on-set 
of Incast events and proactively slows down and limits everybody to a 
constant minimal rate. We also address the recent increase in popularity 
of DCNs adopting the Software Defined Networking (SDN) paradigm which 
differs from the conventional way of networking. To this end, we re-design 
our proposed systems to leverage the powerful control of SDN-based DCNs in 
order to implement SDN-based congestion and traffic control schemes.

Finally, we demonstrate the effectiveness of our proposals by conducting 
simulation and real- testbed experiments on various network topologies. 
Then, we point out possible future directions to be explored before final 
thesis completion.


Date:			Wednesday, 1 March 2017

Time:                  	2:30pm - 4:30pm

Venue:                  Room 4475
                         (lifts 25/26)

Committee Members:	Dr. Brahim Bensaou (Supervisor)
  			Dr. Pan Hui (Chairperson)
 			Prof. Gary Chan
  			Dr. Kai Chen


**** ALL are Welcome ****