1. (MRC) 2 These slides are not approved for public release Resilient high-dimensional datacenter 1 Control Plane: Controllers and Switches

2. (MRC) 2 These slides are not approved for public release Resilient high-dimensional datacenter switching Approach: high-dimensional structure with a switchlet for every compute node Multi-path redundant topology for performance, resilience and security Components are functionally interchangeable High(er) performance through closer processor- network affinity Remap network topology to match program data flow 2 Non-traditional world view!

3. (MRC) 2 Resilient Realtime Data Delivery 3 Datacenter applications suffer from insufficient network isolation. R2D2 implements prioritized, distributed admission control using unmodified, commodity hardware. R2D2 reduces in-network interference latencies by over 300× and performs on- par with, or better than more complex congestion control schemes including DCTCP, ECN, 802.3x, PDQ, and pFabric R2D2

4. (MRC) 2 These slides are not approved for public release 4

5. (MRC) 2 These slides are not approved for public release The quest for scalable network experimentation 5 Fidelity Scalability Reproducibility How to evaluate exciting new ideas on future networking ? o Increasing network sizes o Complex systems (OS + net functions virtualization, SDN) o 10GbE and beyond (100 GbE?) link speeds o Complex application-level behavours Challenges Key properties for a modern network experimentation platform Replicate experiment + results: across different platforms over multiple runs Replicate real-system and application behavior with accuracy Ability to reproduce larger scale experiments, maintaining fidelity

6. (MRC) 2 These slides are not approved for public release Simulation vs Emulation vs Hybrid 6 Reproducibility FidelityScalability

7. (MRC) 2 These slides are not approved for public release Selena Design 7 Experiment description Python API Selena compiler Selena compiler Reproducibility: - XAPI-based Python API, automated experiments Scalability - Time dilation for unmodified guests - scalability tuning knobs (trade time for fidelity) Fidelity - link emulation - realistic OpenFlow switch models - unmodified code execution: real stacks and OS, full POSIX support

8. (MRC) 2 These slides are not approved for public release Fidelity evaluation 8 Star topology Fat-tree topology Execution time: Ns3  175min 24sec Selena  20min Execution time: Ns3  175min 24sec Selena  20min Execution time: Ns3  172min 51sec Selena  40min Execution time: Ns3  172min 51sec Selena  40min

9. (MRC) 2 These slides are not approved for public release Demo Description 9 Fat-Tree topology (K=4) Multi-Layered OpenFlow controller architecture L-2 routing controlled by SDN

10. (MRC) 2 These slides are not approved for public release Current work 10 Make Selena publicly available (open source code): - http://selena-project.github.io Improve scalability - Multi-machine emulation - Optimize guest-2-guest Xen communications Use cases - Evaluation of scenarios with layered OpenFlow controllers - SDN coupling with workload consolidation Publications ACM/IEEE Symposium on Architectures for Networking and Communications Systems (ANCS) 2014 Faithful Reproduction of Network Experiments Dimosthenis Pediaditakis, Charalampos Rotsos, Andrew W. Moore (University of Cambridge)

11. (MRC) 2 These slides are not approved for public release 11

12. (MRC) 2 These slides are not approved for public release Directions and Research Questions Can we exploit the NetFPGA10G platform and Bluespec OpenFlow switch to integrate R2D2 with resilient switchlets? Does datacenter mesh networking improve resilience, performance & power use? Can we architect a hierarchical control system for it? Can we use capabilities to identify communication flow? 12