Automated and Scalable QoS Control - For Network Convergence Wonho Kim (Princeton Univ.) Puneet Sharma, Jeongkeun Lee, Sujata Banerjee, Jean Tourrilhes, Sung-Ju Lee, and Praveen Yalagandula (HP Labs) 1
Motivation • Why do we care about QoS control? − Network convergence − Multi-tenancy networks • Automated QoS control is needed 2 29 April 2010
Network convergence Ethernet Fibre Channel Map- Storage Reduce Infiniband … HPC … Different protocols, adapters, switches, and configuration 3 29 April 2010
Network convergence Map- Storage Reduce HPC … Fewer switches, ports, adapters, cables Converged Enhanced Ethernet (CEE) Reduced power, equipment, cooling cost Data Center Ethernet (DCE) Simpler topology Data Center Bridging (DCB) I/O consolidation Fibre Channel over Ethernet (FCoE) • Cost reduction (single network infrastructure) Unified resource management Fibre Channel over CEE (FCoCEE) • Unified management (single management infrastructure) 4 29 April 2010
Multi-tenancy networks Customer A Customer C Customer B • Serve multiple customers with a single fabric • Better utilization of network infrastructure 5 29 April 2010
Performance isolation Map- Storage Reduce HPC … 6 29 April 2010
Performance isolation Map- Storage Reduce HPC … 7 29 April 2010
Performance isolation Map- Storage Reduce HPC … 8 29 April 2010
Performance isolation Customer C Customer A Map- Virtualized Servers Storage Reduce Variable Workloads HPC … Bugs, malicious attack Customer B • Need virtual network slices • Need fine-grained performance isolation 9 29 April 2010
Goal • Enables performance isolation with QoS control Slice 1 Slice n … Physical network fabric QoS configuration High-level Specifications 10 29 April 2010
Good news • Most commodity switches have QoS knobs − rate limiter − priority queues − schedulers • Single network domain − datacenters, enterprise networks, … − free from Layer-8 issues (billing, collaborations, …) − fine-grained control becomes feasible 11 29 April 2010
Challenges • Coarse-grained QoS knobs − designed for distributed management − class-based − no e2e performance • Manual configuration Storage VoIP − no standards for classifiers Traffic Fibre Channel − error-prone CEO/CTO Email Managers − static (not adaptive) Web Interns 12 29 April 2010
Our Solution: OpenFlow QoS Controller Virtual Slice 1 Virtual Slice n … Physical network fabric QoS configuration (rate limiter, priority queues, …) Problem Manual & Static setting High-level • Automated configuration QoS Controller Specifications • Fine-grained flow management • Adaptive to dynamic workloads 13 29 April 2010
Overview of OpenFlow QoS controller Flow Dest Src rate limit prio prio prio prio OpenFlow APIs Rate limiters Network-wide Adaptive QoS Controller Optimization Aggregator Queue mappings Shortest Span First … Topology, Nodes, End-to-End Resource, Flows, … performance models 14 14 29 April 2010
Adaptive aggregation Flow specs Slice specs flow 1 Peak rate: 400 Mbps Customer DB Storage Delay bound: 10 ms src IP: X.X.X.X, port: 9551 Aggregate: False flow 2 Type: IP (UDP), … flow 3 flow 3 Peak rate: 100 Mbps Video Employee DB Delay bound: 100 ms flow 4 Aggregate: False flow 5 VoD flow 6 System backup flow 7 Aggregate: True Backup flow 8 Log Archive 15 29 April 2010
Available QoS Knobs (Priority queue) highest short delay Output port long delay lowest 16 29 April 2010
Available QoS Knobs (Rate limiter) highest Peak rate Output port lowest 17 29 April 2010
OpenFlow QoS APIs Flow 1 H/W rate limiter 1 Flow 2 Rate limiter Flow 3 H/W rate limiter 2 Flow 4 Priority queue mapping Switch 1 Switch 2 Switch 3 • Extension of OpenFlow specification • Expose QoS capability in switches 18 18 29 April 2010
OpenFlow QoS APIs • With OpenFlow flow control − fine-grained control of flows − automated flow management • With OpenFlow QoS APIs − uniform control of QoS knobs − configure QoS for individual (or aggregate) flows 19 29 April 2010
Admission Control • Input − new flow arrival event − performance requirements (peak rate, e2e delay) − database for the current network state − end-to-end performance model − end-to-end performance model • Output − admission control result (accept/reject) − priority queue assignment, rate limiter settings − path selection 20 29 April 2010
Admission Control • Two conditions should be satisfied − satisfy f ’s performance requirement − not violate existing flows in the networks 21 29 April 2010
Difficulties in queue assignment highest (8/10) (10/10) f1 (8/10) (10/10) f2 lowest Switch A Switch B Switch C f We should consider interactions between • flows in a switch • flows in multiple switches 22 29 April 2010
Admission control heuristic • Goal − increase the ratio of admitted flows − lower the complexities in queue allocation • Shortest Span First (SSF) • Basic ideas − estimate affordable options for a flow − try first switches more likely to reject flow 23 29 April 2010
Highest level & Lowest level high Shorter delay bound Highest level for f More likely to violate other flows f Longer delay bound Lowest level for f More likely to violate flow f low Switch • Highest level: not violate existing flows • Lowest level: not violate the new flow • Span: available options for f 24 29 April 2010
Shortest Span First (SSF) high src dest f Switch A low Switch B Switch C • Step 1: compute highest & lowest levels independently 25 29 April 2010
Shortest Span First (SSF) high src dest f Switch A low Switch B Switch C • Step 2: sort switches in order of the span 26 29 April 2010
Shortest Span First (SSF) high src dest f Switch C low Switch A Switch B • Step 2: sort switches in order of the span 27 29 April 2010
Shortest Span First (SSF) high src dest f Switch C low Switch A Switch B • Step 3: try highest level at each hop − try first a switch more likely to reject flow 28 29 April 2010
Implementataion • QoS APIs implemented on − hardware switch (HP ProCurve 5406zl) − software switch (Open vSwitch) • QoS Controller implemented on top of NOX − open-source OpenFlow controller − http://noxrepo.org • QoS Controller web interface 29 29 April 2010
Prototype QoS Host A Host B Controller OpenFlow Output 14 Switch A Priority 7 OpenFlow HTTP OpenFlow Output 21 Switch C Switch B Priority 7 Peak rate 400 Mbps 400 Delay bound 10 ms Host C Host D 30 29 April 2010
Evaluation • Traffic generation − generate 3 guaranteed flows from emulated services (UDP) − generate cross traffic (UDP, TCP) • Disable/Enable QoS controller • Measured throughput and packet loss in testbeds 31 29 April 2010
Throughput with UDP cross traffic Generate cross traffic Flow name Route (queue assignment) Customer DB H3 – S3(8) – S1(8) – H1 Employee DB H4 – S3(8) – S1(8) – H2 VoD H3 – S3(7) – S1(7) – H1 System Backup H4 – S3(1) – S1(1) – H2 Enable QoS Controller QoS controller protects guaranteed flows in congestion 32 29 April 2010
Packet loss with TCP cross traffic 30 TCP 10 TCP 1 TCP Enable QoS Controller QoS control is needed even when most traffic in network is TCP 33 29 April 2010
Future works • Evaluations − effectiveness of admission control heuristics (ratio of admitted flows) − compare with offline optimal assignment − simulations on a variety of datacenter networks (e.g., − simulations on a variety of datacenter networks (e.g., Hierarchical, FatTree, …) • Deployment − extend deployment to large networks − test with mixture of services 34 29 April 2010
Conclusion • Single integrated network fabric is desirable • We need fine-grained automated QoS control • Contributions − Design & Implement OpenFlow QoS APIs − QoS controller: automated QoS control for network slicing 35 29 April 2010
Thank you 36 29 April 2010
Recommend
More recommend
