Maslows hierarchy of network programming and the unmet

Maslow’s hierarchy of network programming and the unmet needs Ratul Mahajan Microsoft Research

How we got here

How we got here Programmability was not a goal!

Why we can escape our fate now Many new networks r o opp y t i tun Bigger networks pr d e e n g n i s s e

Microsoft Azure: Newer and bigger networks 2010 2015 Data centers A few 100+ Network devices 1000 s 100 s of 1, 000 s Network capacity 10 s of Tbps Pbps

Complexity Hypothesis on complexity of network management Network size

Why we can escape our fate now Many new networks r o opp y t i tun Bigger networks pr d e e n g n i s s e Standardized abstractions key r e l b ena

Maslow’s hierarchy for network programming Efficiency Policy-compliant forwarding Bootstrapped, healthy devices Human Network

Maslow’s hierarchy for network programming Merlin, ? ? Unmet needs Efficiency {Fren, Pyr, Kin, . . }etic, Net. KAT, Propane ? ? Policy-compliant forwarding Bootstrapped, healthy devices Network

Enabling abstractions for network programming ? ? Unmet needs Efficiency Open. Flow, BGP Policy-compliant forwarding ? ? Bootstrapped, healthy devices Network

Past: Management apps interacted with devices Switch upgrade Link corruption mitigation Elastic scaling Conflicts among apps Hard to guarantee SLAs Hard to develop applications

Anatomy of a management app Switch upgrade Link corruption mitigation Elastic scaling App logic Mon Upd itor ate

Statesman: Network state management service App logic Network state database Monitor [A network-state management service, SIGCOMM 2014] Update Power on/off Firmware version Link state Tunnels …. .

Statesman: Network state management service App logic Observed state Proposed state Monitor Update Target state

State databases emerging at device level too SAI and SONi. C (OCP) https: //github. com/opencomputeproject/SAI https: //github. com/Azure/SONi. C Switch state service

What is missing from state databases High-level languages to read and write network state – Program network as a whole, given operator goals – Statesman, So. NIC are low-level • E. g. , read/write individual state variables

Enabling abstractions for network programming ? ? Unmet needs Efficiency Open. Flow, BGP Network state ? ? databases Policy-compliant forwarding Bootstrapped, healthy devices Network

Inter-DC WAN: A critical, expensive resource Seattle Seoul Dublin New York Barcelona Hong Kong Los Angeles Miami But highly inefficient

Cause of inefficiency: Lack of coordination

SWAN: Software-driven WAN Goals Key design elements Highly efficient WAN Flexible sharing policies Coordinate across services Centralize resource allocation [Achieving high utilization with software-driven WAN, SIGCOMM 2013]

SWAN overview SWAN controller Traffic demand BW allocation Service broker Rate limiting Service hosts Topology, traffic Network config. Network agent WAN

Throughput (relative to optimal) SWAN comes close to optimal SWAN w/o rate control MPLS TE

Resource allocation in SWAN Policies Priority classes Fairness Resources Link capacities Switch memory Enabling abstractions Tunnels Priority queues Rate limiters ECN marking

Enabling abstractions for network programming Tunnels, priority queues, rate ? ? limiters Open. Flow, BGP Network state databases Unmet needs Efficiency Policy-compliant forwarding Bootstrapped, healthy devices Network

Trustworthy validation: A cross-cutting need Programmatic control often gets gated on validation Much progress in recent years but challenges remain Update Where to validate? Model for design Model for validation Reason in face of dynamics Permissive validation models

Opportunistic time to make networks easier to program Let us focus also on unmet needs and on trustworthy validation Deployed systems have validated enabling abstractions SWAN Unmet needs Efficiency Tunnels, priority queues, rate limiters Policy-compliant forwarding Statesman, SAI Bootstrapped, healthy devices Network state databases