ECE Seminar: Efficient Traffic Exchange at Internet Exchange Points
Efficient Traffic Exchange at Internet Exchange Points
Ibrahim Alam, Rensselaer Polytechnic Institute
Friday, November 22 at 12:00 PM
17 HLH, Rm 328 or Zoom (https://yale.zoom.us/j/93572530213)
Hosted by: Leandros Tassiulas
Abstract:
End users rely on Internet Service Providers (ISPs) for access to the Internet. Given the limited coverage of these ISPs within specific geographic regions, the establishment of interconnectivity between ISPs is crucial for global connectivity. Internet Exchange Points (IXPs) are equipped with extensive network switches that enable ISPs to establish peering connections facilitating direct traffic exchange. This work presents a series of proposed policies aimed at improving the efficiency of the Internet traffic exchange of ISPs at IXPs. Firstly, we investigate two pricing policies of IXPs. We show that IXPs should implement a constant pricing policy when the internet demand (curve) maintains a specific pattern (or shape) and the policy will ensure good utility and revenue simultaneously. On the contrary, if the internet demand curve is dynamic, i.e., not very stable over time, the IXP(s) may adopt a proportional pricing policy. Secondly, we analyze the port capacity purchase (PCP) problem for ISPs at an IXP. When making port capacity purchase decisions unilaterally and traffic exchange decisions bilaterally, the ISPs should focus on maximizing their utility. Due to the nature of the market, this utility maximization of each ISP cannot hurt (much) the total maximum utility of the market. Thirdly, we study the automation of the two steps of the peering decision process: i) if two ISPs should peer, and ii) where they should peer. We found that ISP peering partner selection (with which ISP to peer) is dependent mostly on the characteristics of the respective ISP pair and does not depend on the whole market. Finding the best ISP peering location(s) is a complicated problem. A set of good solutions with peering locations can be recommended but is computationally expensive.
At the end of the talk, we will also discuss some results on i) the Combined Algorithm Selection and Hyper-Parameter (HP) Optimization (CASH) problem in the Federated Learning (FL) setting, and ii) the utilization of fine-tuned Large Language Models (LLMs) to perform data-scarce tasks, done in collaboration with IBM.
Bio:
Ibrahim Alam is a Ph.D. candidate in the ECSE department at Rensselaer Polytechnic Institute; expected to graduate in December 2024. He was awarded the Founders Award of Excellence from RPI in October 2024 for his research. He received his B.Sc. and M.Sc. degrees from the Bangladesh University of Engineering and Technology (BUET) in 2014 and 2017, respectively. Following his bachelor's degree, he joined Huawei in 2014, where he worked for a year. He then transitioned to academia as a Lecturer in college in 2015. Ibrahim started his Ph.D. in 2019 and has been involved in diverse research projects, including efficient traffic exchange using game theory, AutoML in federated learning, and ensemble methods for large language models in data-scarce domains.