Efficiency and Resilience in Data Centers
- with Dr. Ihsan Ayyub Qazi, Dr. Zartash Afzal Uzmi and Dr. Fahad R. Dogar
In Data Centers there is always a trade-off in having an efficient tenant allocation versus providing resilient service to the tenants. Packing virtual machines under same Top of the Rack (TOR) Switch or under same physical machines consumes relatively little bandwidth. However, spreading virtual machines across the Data Center gives a fair amount of resilience while consuming a lot of bandwidth. Our work involves proposing and evaluating different schemes to provide excellent fault tolerance while maintaining reasonable bandwidth consumption.
Forum Thread Length Prediction [Paper]
- with Dr. Muhammad Fareed Zaffar
Online social media allows consumers to engage with each other and to create, share, discuss and modify user-generated content in a highly interactive way. Social media platforms have therefore become critical for companies trying to gauge the pulse of consumers, help identify issues faster, receive immediate feedback on products and offering. In our work we aimed to propose and evaluate a classification based framework to predict thread lengths in online discussion forums in order to identify potential topics that may be of interest to a particular online community. Our work was accepted at the PACIS 2015 conference.
Minimizing Statistics Collection in SDN-enabled Data Centers [Technical Report]
- with Dr. Ihsan Ayyub Qazi and Dr. Zartash Afzal Uzmi
Software Defined Networking is a paradigm that separates the data plane and the control plane. Openflow has become the standard for communication between the two planes. It allows us to gather statistics and let network applications use these statistics. However, regular pinging for statistics and requesting from multiple switches can lead to a large amount of traffic in the control plane. Our work aimed at providing a method to reduce overhead. Included is a proof that while it pings less frequently and fewer switches, it provides highly accurate information.
On the Effectiveness of CoDel in Data Centers [Technical Report]
- with Dr. Ihsan Ayyub Qazi
CoDel (Controlling Delay) is an AQM scheme proposed for wide-area networks to address bufferbloat; a condition in which excessive buffering leads to high latency. Previous work was evaluated in wide-area network but no testing had been done in Data Center environments. We rigorously evaluated the effectiveness of CoDel compared to RED and simple DropTail in different scenarios pertaining to Data Centers.
Minimizing Flow Completion Times in Data Centers [Paper]
- with Dr. Ihsan Ayyub Qazi and Dr. Zartash Afzal Uzmi
For provisioning large-scale online applications data centers face extreme challenges in providing low latency for short flows (that result from end-user actions) and high throughput for background flows (that are needed to maintain data consistency and structure across massively distributed systems). We proposed L^2DCT, a data center transport protocol that aims at reducing flow completion times for short flows by approximating the Least Attained Service (LAS) schedule discipline without requiring changes in application software or router hardware. Our work was accepted at the IEEE INFOCOM 2013 conference.
- with Dr. Ihsan Ayyub Qazi, Dr. Zartash Afzal Uzmi and Dr. Fahad R. Dogar
In Data Centers there is always a trade-off in having an efficient tenant allocation versus providing resilient service to the tenants. Packing virtual machines under same Top of the Rack (TOR) Switch or under same physical machines consumes relatively little bandwidth. However, spreading virtual machines across the Data Center gives a fair amount of resilience while consuming a lot of bandwidth. Our work involves proposing and evaluating different schemes to provide excellent fault tolerance while maintaining reasonable bandwidth consumption.
Forum Thread Length Prediction [Paper]
- with Dr. Muhammad Fareed Zaffar
Online social media allows consumers to engage with each other and to create, share, discuss and modify user-generated content in a highly interactive way. Social media platforms have therefore become critical for companies trying to gauge the pulse of consumers, help identify issues faster, receive immediate feedback on products and offering. In our work we aimed to propose and evaluate a classification based framework to predict thread lengths in online discussion forums in order to identify potential topics that may be of interest to a particular online community. Our work was accepted at the PACIS 2015 conference.
Minimizing Statistics Collection in SDN-enabled Data Centers [Technical Report]
- with Dr. Ihsan Ayyub Qazi and Dr. Zartash Afzal Uzmi
Software Defined Networking is a paradigm that separates the data plane and the control plane. Openflow has become the standard for communication between the two planes. It allows us to gather statistics and let network applications use these statistics. However, regular pinging for statistics and requesting from multiple switches can lead to a large amount of traffic in the control plane. Our work aimed at providing a method to reduce overhead. Included is a proof that while it pings less frequently and fewer switches, it provides highly accurate information.
On the Effectiveness of CoDel in Data Centers [Technical Report]
- with Dr. Ihsan Ayyub Qazi
CoDel (Controlling Delay) is an AQM scheme proposed for wide-area networks to address bufferbloat; a condition in which excessive buffering leads to high latency. Previous work was evaluated in wide-area network but no testing had been done in Data Center environments. We rigorously evaluated the effectiveness of CoDel compared to RED and simple DropTail in different scenarios pertaining to Data Centers.
Minimizing Flow Completion Times in Data Centers [Paper]
- with Dr. Ihsan Ayyub Qazi and Dr. Zartash Afzal Uzmi
For provisioning large-scale online applications data centers face extreme challenges in providing low latency for short flows (that result from end-user actions) and high throughput for background flows (that are needed to maintain data consistency and structure across massively distributed systems). We proposed L^2DCT, a data center transport protocol that aims at reducing flow completion times for short flows by approximating the Least Attained Service (LAS) schedule discipline without requiring changes in application software or router hardware. Our work was accepted at the IEEE INFOCOM 2013 conference.