Mirror Mirror on the Ceiling:Flexible Wireless Links for Data Centers

Presenter: Lu Gong

About Authors

About Authors: Xia Zhou• PhD candidate, UCSB• Fields:

Networks & Communications, Mathematics, Microbiology

About Authors: Zengbin Zhang• PhD candidate, UCSB• Fields: Wireless Systems and Networking, Mobile

Computing and Distributed Systems

About Authors: Yibo Zhu• PhD candidate, UCSB• Fields: Data Center Networks, Mobile

Networks and Online Social Networks

Problems with any wired network topology

• Any large-scale network consists of multiple stages→ nr of fibers/wires are doubled/tripled

• Distribute fixed amount of fibers to every rack→ fibers are over deployed for the worst case

• Once deployed, very hard to modify

Our goal• Focus on a subset of applications

– that do not require non-blocking all-to-all communication– exclude high-end datacenter computing

• We hope to create a new primitive– high-throughput, beamforming wireless links in the 60GHz band

Existing works

• Signal leakage→ limits the concurrent active links

• Line-of-sight requirement→ limits the effective range of links

Properties of 60GHz band wireless links

• 7GHz spectrum → multi-Gbps bandwidth• High frequency → small interference• Able to use beamforming to enhance link rate

and further suppress interference• 5mm wavelength → any object larger than

2.5mm can block/reflect signal

Beamforming

• A physical layer technique to concentrate transmission energy in a specific direction

Testbed of link blockage

Testbed of radio interference

3D Beamforming

• Components:– Beamforming Radios– Ceiling Reflectors– Electromagnetic Absorbers

• Prevent local reflections and scattering

Testbed of 3D Beamforming

Microbenchmark: Validate Physical Properties

Microbenchmark: Radios per Rack

Microbenchmark: Sensitivity to Hardware

Scheduling: Goal & Challenges

• Goal– Maximize efficiency– Minimize wireless interference

• Challenges– Require accurate interference model (accumulate interference)– Handle short-lived traffic burst → must be online– Account for antenna rotation delay (0.01s~1s)

Scheduling: Design

• Conflict-Degree based Greedy Scheduling– Goal → Minimize job completion time– Graph coloring problem

• Color: 60GHz frequency channels & time slots– Link preemption or not?– Minimize antenna rotation overhead

Evaluation: Addressing Traffic Hotspots

• Does adding 3D beamforming links to existing wired networks significantly increase available bandwidth for hotspots?

• How significant are the benefits of 3D beamforming over 2D beamforming, and where are they most visible?

• Will antenna rotation delay of today’s rotators be a performance bottleneck for 3D beamforming?

Evaluation: Flow Completion Time

Future Work

• Routing• Traffic management• Priority-based scheduler• Wired/wireless co-scheduling

Related Work

• Address traffic congestion– Network architecture design & traffic scheduling– Modeling network traffic characteristics

• 60GHz wireless technology• Optical circuit switching

Thank you for listening