Publications

Conference Publications:

  • Syed Mohammad Irteza, Hafiz Mohsin Bashir, Talal Anwar, Ihsan Ayyub Qazi, and Fahad Rafique Dogar. "Load Balancing Over Symmetric Virtual Topologies", in Proceedings of IEEE INFOCOM 2017, Atlanta, GA, USA, May 2017
  • Ali Munir, Ghufran Baig, Syed Mohammad Irteza, Ihsan Ayyub Qazi, Fahad Rafique Dogar, and Alex Liu. "Friends, not Foes -- Synthesizing Existing Transport Strategies for Data Center Networks", in Proceedings of ACM SIGCOMM 2014, Chicago, IL, USA, August 2014
  • Syed Mohammad Irteza, Adnan Ahmed, Sana Farrukh, Babar Naveed Memon, and Ihsan Ayyub Qazi. "On The Coexistence of Transport Protocols in Data Centers", in Proceedings of IEEE ICC 2014, Sydney, Australia, June 2014


Journal Publications:

  • Syed Mohammad Irteza, Hafiz Mohsin Bashir, Talal Anwar, Ihsan Ayyub Qazi, and Fahad Rafique Dogar. "Efficient Load Balancing Over Asymmetric Datacenter Topologies", in Computer Communications, Elsevier, May 2018
  • Ali Munir, Ghufran Baig, Syed Mohammad Irteza, Ihsan Ayyub Qazi, Fahad Rafique Dogar, and Alex Liu. "PASE: Synthesizing Existing Transport Strategies for Near-Optimal Data Center Transport", IEEE/ACM Transactions on Networking, September 2016 (Issue 99)


PhD Thesis:

  • "Resilient Network Load Balancing for Datacenters" - December 2018, Department of Computer Science, SBASSE, LUMS
    • Advisor: Dr. Ihsan Ayyub Qazi

Abstract: Datacenters (DCs) provide the backbone for online services such as web search and social networks. To deliver high bandwidth to such services, modern datacenters are often structured as multi-rooted tree topologies, resulting in multiple paths between every pair of servers. To optimally utilize the available bandwidth, an efficient network load balancing scheme is a critical component. Studies show that datacenters are prone to link failures and switch malfunctions, and this results in congestion asymmetry across network paths. ECMP is a widely deployed load balancing scheme in switches today, which hashes each flow onto one specific path. This leads to hash collisions for elephant flows. On the other side of the spectrum, some schemes spray packets across multiple paths, and while their performance is near-optimal under symmetric conditions, their performance degrades significantly under asymmetry, due to packet reordering. In this thesis, we explore how per-packet load balancing can be made efficient in the presence of asymmetry.

The key contribution of this thesis is the development of per-packet datacenter network load balancing schemes that perform well under asymmetry. Our work is based on the observation that if each flow is provided with a topology that exhibits symmetry in congestion across all its paths, then the flow is able to fully benefit from the bisection bandwidth without facing packet reordering. To this end, we propose SAPS, i.e., Symmetric and Adaptive Packet Spraying. SAPS leverages flow and group tables within OpenFlow-enabled switches to create symmetric virtual topologies (SVTs), with each SVT consisting of paths with similar capacity. SAPS maps each flow to a specific SVT, with elephant flows mapped to larger bandwidth SVTs, and mice flows mapped probabilistically to any SVT. Experiments conducted involving large-scale packet-level simulations and real testbed settings indicate that SAPS improves upon the performance of several existing load balancing schemes such as ECMP and weighted packet spraying, under various failure scenarios. Under certain scenarios, it also improves upon state-of-the-art schemes such as FlowBender and CONGA.

OpenFlow-enabled switches provide the ability to rate limit groups of flows at the switch. By leveraging the rate limiting feature within OpenFlow-enabled switches, we can construct an SVT by extracting a specific symmetric chunk (or slice) of residual capacity from various paths, rather than the all or nothing path inclusion mechanism in SAPS. Whereas SAPS works well when asymmetry affects a few paths, its performance is adversely affected under higher levels of asymmetry. North-south traffic can lead to a greater number of paths being affected by asymmetry. Rate limiting enables the construction of a set of SVTs that contains the largest possible single SVT. We observe that under greater asymmetry, having an SVT set with the maximum possible SVT significantly enhances performance. This forms the basis of our proposed scheme, Slices, which outperforms other network load balancing schemes (including SAPS) when most paths face asymmetry. Both schemes (SAPS and Slices) are deployable on commonly available OpenFlow-enabled switches.