PhD Thesis Defense: Jorge Murillo, Carbon-Aware Spatial Shifting for Internet-Scale Interactive Services
Content
Speaker:
Abstract:
Demand for Internet-scale services such as social media and video streaming has grown rapidly over the past fifteen years. These services reach billions of users worldwide and must serve millions of latency-sensitive requests every minute. To do so, they rely on a vast, geographically distributed infrastructure of datacenters and networks, deployed across many regions to reduce the latency of serving requests. This infrastructure consumes enormous amounts of electricity. US datacenters alone are projected to consume 6.7–12% of national electricity by 2028, which makes Internet-scale services a significant source of greenhouse gas emissions.
Yet the same distributed infrastructure that enables these services can be leveraged to reduce their own emissions. Because datacenters span regions whose electricity grids differ widely in carbon intensity, requests can be spatially shifted to datacenters in locations powered by cleaner energy. However, shifting requests can increase latency and energy costs, and capacity and peak-demand constraints limit how much load can move. Moreover, previous work on carbon-aware shifting has typically ignored the emissions of the network, focusing only on servers as the main source of emissions.
This thesis develops spatial shifting techniques that reduce the carbon emissions of Internet-scale services while preserving their performance. First, I present CDN-Shifter, a multi-objective optimization approach for content delivery networks. It combines load shifting with capacity shifting under latency and energy-cost constraints. On Akamai CDN traces, it reduces emissions by 35.5% in the US with only a 60 ms latency increase, and by up to 57% with capacity shifting. Second, in Go with the Flow, I study video streaming, which accounts for over half of Internet traffic. Using an analytical model that attributes emissions to the server, the network path and the playback device, I show that network-agnostic shifting can yield lower than expected, or even negative, savings. Third, I present NG-Stream, which incorporates per-path network emissions, capacity and peak-demand constraints into a global optimization framework for video delivery. On Akamai traces, NG-Stream reduces emissions by 31% in the US and 69% in Europe, compared with 7.4% and 46.3% for network-agnostic shifting.
Together, these results show that spatial shifting can substantially decarbonize Internet-scale services while preserving their performance and controlling operating costs.
Advisor:
Prashant Shenoy