All seminars will take place on Fridays at 11 a.m. in DBH 6011. Check seminar details below.
Sharad Agarwal
Microsoft
December 2, 2022
11:00am - 12:00pm
Title:
Improving the Reliability of 5G Traffic Traversing Hyperscalar Networks
Abstract:
I will begin by describing why 5G networks are radically different from past Gs and why cloud hyperscalars play an important role in 5G. I will tell you why I am excited by the many opportunities computer science researchers have to innovate in solving the requirements of 5G networks. Challenges such as high reliability, differentiated QoS, security, and efficient use of compute span many disciplines including networking, wireless communication, mobile systems, and operating systems.
I will then focus on one specific problem -- improving the reliability of 5G traffic traversing hyperscalar networks. One class of 5G traffic requires 99.999% reliability, which is tough to achieve end-to-end. Peering links at the boundary of hyperscalar networks are a significant source of unreliability. Existing techniques can manage egress traffic on those links during outages or congestion, but managing ingress traffic has remained a hard technical challenge.
We built a learning system called TIPSY (Traffic Ingress Prediction SYstem) that predicts on which peering link customer traffic will arrive, even under route withdrawals or link outages. Its accuracy is 97.9% for all ingress traffic, and 76.4% specifically during outages. We are running it for the Azure wide-area network to mitigate traffic congestion. It has demonstrated accuracy in the post-mortem of multiple incidents. This work was accepted at ACM SIGCOMM 2022 and is with Michael Markovitch, Rodrigo Fonseca, Ryan Beckett, Jenny Zhang, Irena Atov, and Somesh Chaturmohta.
I will then focus on one specific problem -- improving the reliability of 5G traffic traversing hyperscalar networks. One class of 5G traffic requires 99.999% reliability, which is tough to achieve end-to-end. Peering links at the boundary of hyperscalar networks are a significant source of unreliability. Existing techniques can manage egress traffic on those links during outages or congestion, but managing ingress traffic has remained a hard technical challenge.
We built a learning system called TIPSY (Traffic Ingress Prediction SYstem) that predicts on which peering link customer traffic will arrive, even under route withdrawals or link outages. Its accuracy is 97.9% for all ingress traffic, and 76.4% specifically during outages. We are running it for the Azure wide-area network to mitigate traffic congestion. It has demonstrated accuracy in the post-mortem of multiple incidents. This work was accepted at ACM SIGCOMM 2022 and is with Michael Markovitch, Rodrigo Fonseca, Ryan Beckett, Jenny Zhang, Irena Atov, and Somesh Chaturmohta.
Speaker Bio:
Sharad Agarwal is a Senior Principal Researcher at Microsoft. In addition to 40 publications, his research has impacted Azure, Windows, and Windows Phone. He has helped design and deploy a large cloud service on Azure that has served over 21 million users. He is an ACM Distinguished Member. In his spare time, he is an active angel investor with https://www.grubstakes.vc/ in the Seattle startup ecosystem.