Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about datacenter network optimization through this 18-minute conference presentation from NSDI '25 that examines network hotspots in Google's datacenter infrastructure. Discover how researchers from Google, Harvard University, University of Southern California, and Columbia University identified and addressed persistent network bottlenecks that occur at Top-of-Rack (ToR) switches and can last for hours. Explore the root causes of these hotspots, including bandwidth demand-supply imbalances from network-intensive services and infrastructure upgrade mismatches where compute and storage capabilities outpace network bandwidth improvements. Understand how traditional bandwidth-independent task and data placement by datacenter schedulers contributes to performance degradation, with some distributed applications experiencing over 2× latency increases during hotspot conditions. Examine the practical solutions implemented at Google, including hotspot-aware task placement in cluster schedulers that reduced hot ToRs by 90%, and hotspot-aware data placement in distributed file systems that decreased p95 network latency by more than 50%. Gain insights into why strategic placement of tasks and data under ToRs with higher available bandwidth proves more effective than relying solely on congestion control, load balancing, and traffic engineering for achieving consistent datacenter performance.
