Friends, not Foes – Synthesizing Existing Transport Strategies for Data Center Networks Ali Munir Michigan State University Michigan State University Ghufran Baig, Syed M. Irteza, Ihsan A. Qazi, Alex X. Liu, Fahad R. Dogar
Data Center (DC) Applications Distributed applications Map- Map- Search Mail Components interact via the Reduce Reduce network e.g., a bing search query touches HPC Monitoring > 100 machines Network impacts performance “10% of search responses observe 1 to 14 ms of network queuing delay” [ DCTCP, SIGCOMM 10] Image source: http://cdn.slashgear.com/wp-content/uploads/2012/10/google-datacenter-tech-13.jpg
DC Network Resource Allocation Fair Sharing Equal bandwidth sharing among jobs [TCP, DCTCP] – Increases completion time for everyone – Traditional “fairness” metrics less relevant QoS Aware Prioritize some jobs over other jobs (Priority Scheduling) – Minimize flow completion times [pFabric, L 2 DCT] – Meet flow deadlines [D 3 , D 2 TCP]
DC Transports pFabric DCTCP SIGCOMM’13 D 3 SIGCOMM’10 L 2 DCT SIGCOMM’11 INFOCOM’13 PDQ D 2 TCP SIGCOMM’12 SIGCOMM’12
DC Transports pFabric DCTCP SIGCOMM’13 Near Optimal but not D 3 SIGCOMM’10 L 2 DCT SIGCOMM’11 Deployment Friendly INFOCOM’13 PDQ (Changes in data plane) D 2 TCP SIGCOMM’12 SIGCOMM’12
DC Transports pFabric DCTCP SIGCOMM’13 Near Optimal but not D 3 SIGCOMM’10 Deployment Friendly but L 2 DCT SIGCOMM’11 Deployment Friendly Suboptimal INFOCOM’13 PDQ (Changes in data plane) D 2 TCP SIGCOMM’12 SIGCOMM’12
DC Transports pFabric DCTCP SIGCOMM’13 Near Optimal but not D 3 SIGCOMM’10 Deployment Friendly but L 2 DCT SIGCOMM’11 Deployment Friendly Suboptimal INFOCOM’13 PDQ (Changes in data plane) D 2 TCP SIGCOMM’12 SIGCOMM’12 Step back and ask How can we design a deployment friendly and near optimal data center transport while leveraging the insights offered by existing proposals?
DC Transports pFabric DCTCP SIGCOMM’13 Near Optimal but not D 3 SIGCOMM’10 Deployment Friendly but L 2 DCT SIGCOMM’11 Deployment Friendly Suboptimal INFOCOM’13 PDQ (Changes in data plane) D 2 TCP SIGCOMM’12 SIGCOMM’12 Step back and ask How can we design a deployment friendly and near PASE optimal data center transport while leveraging the insights offered by existing proposals?
Rest of the Talk … DC Transport Strategies PASE Design Evaluation
Rest of the Talk … DC Transport Strategies PASE Design Evaluation
DC Transport Strategies Self-adjusting endpoints e.g., TCP, DCTCP, L 2 DCT – senders make independent decisions and adjust rate by themselves Arbitration e.g., D 3 , PDQ – a common network entity (e.g., a switch) allocates rates to each flow In-network prioritization e.g., pFabric – switches schedule and drop packets based on the packet priority
DC Transport Strategies Self-adjusting endpoints e.g., TCP, DCTCP, L 2 DCT – senders make independent decisions and adjust rate by themselves Arbitration e.g., D 3 , PDQ Existing DC transport proposals use Existing DC transport proposals use – a common network entity (e.g., a switch) allocates only one of these strategies rates to each flow In-network prioritization e.g., pFabric – switches schedule and drop packets based on the packet priority
Transport Strategies in Isolation Transport Example Pros Cons Strategy Self- DCTCP, Adjusting D 2 TCP, Endpoints L 2 DCT Arbitration PDQ, D 3 In-network pFabric Prioritization
Transport Strategies in Isolation Transport Transport Example Example Pros Pros Cons Cons Strategy Strategy Self- Self- DCTCP, DCTCP, No strict priority Adjusting Adjusting D 2 TCP, D 2 TCP, Ease of deployment scheduling Endpoints Endpoints L 2 DCT L 2 DCT Arbitration Arbitration PDQ, D 3 PDQ, D 3 In-network In-network pFabric pFabric Prioritization Prioritization
Transport Strategies in Isolation Transport Transport Transport Example Example Example Pros Pros Pros Cons Cons Cons Strategy Strategy Strategy Self- Self- Self- DCTCP, DCTCP, DCTCP, No strict priority No strict priority Adjusting Adjusting Adjusting D 2 TCP, D 2 TCP, D 2 TCP, Ease of deployment Ease of deployment scheduling scheduling Endpoints Endpoints Endpoints L 2 DCT L 2 DCT L 2 DCT o High flow switching o High flow switching Strict priority overhead Arbitration Arbitration Arbitration PDQ, D 3 PDQ, D 3 PDQ, D 3 scheduling o Hard to compute precise rates In-network In-network In-network pFabric pFabric pFabric Prioritization Prioritization Prioritization
Transport Strategies in Isolation Transport Transport Transport Transport Example Example Example Example Pros Pros Pros Pros Cons Cons Cons Cons Strategy Strategy Strategy Strategy Self- Self- Self- Self- DCTCP, DCTCP, DCTCP, DCTCP, No strict priority No strict priority No strict priority Adjusting Adjusting Adjusting Adjusting D 2 TCP, D 2 TCP, D 2 TCP, D 2 TCP, Ease of deployment Ease of deployment Ease of deployment scheduling scheduling scheduling Endpoints Endpoints Endpoints Endpoints L 2 DCT L 2 DCT L 2 DCT L 2 DCT o High flow switching o High flow switching o High flow switching o High flow switching Strict priority Strict priority overhead overhead Arbitration Arbitration Arbitration Arbitration PDQ, D 3 PDQ, D 3 PDQ, D 3 PDQ, D 3 scheduling scheduling o Hard to compute o Hard to compute precise rates precise rates o Switch-local decisions In-network In-network In-network In-network Low flow switching o Limited # of priority pFabric pFabric pFabric pFabric Prioritization Prioritization Prioritization Prioritization overhead queues
Transport Strategies in Unison Transport Transport Transport Transport Example Example Example Example Pros Pros Pros Pros Cons Cons Cons Cons Strategy Strategy Strategy Strategy Self- Self- Self- Self- DCTCP, DCTCP, DCTCP, DCTCP, No strict priority No strict priority No strict priority Adjusting Adjusting Adjusting Adjusting D 2 TCP, D 2 TCP, D 2 TCP, D 2 TCP, Ease of deployment Ease of deployment Ease of deployment scheduling scheduling scheduling Endpoints Endpoints Endpoints Endpoints L 2 DCT L 2 DCT L 2 DCT L 2 DCT o High flow switching o High flow switching o High flow switching o High flow switching Strict priority Strict priority overhead overhead Arbitration Arbitration Arbitration Arbitration PDQ, D 3 PDQ, D 3 PDQ, D 3 PDQ, D 3 scheduling scheduling o Hard to compute o Hard to compute precise rates precise rates o Switch-local decisions In-network In-network In-network In-network Low flow switching o Limited # of priority pFabric pFabric pFabric pFabric Prioritization Prioritization Prioritization Prioritization overhead queues
Transport Strategies in Unison In-network Prioritization Alone Limited # of queues More # of flows (priorities) High Priority High Priority Flows 1 2 3 4 Low Priority
Transport Strategies in Unison In-network Prioritization Alone Flow Multiplexing Limited # of queues Limited performance gains! More # of flows (priorities) High Priority High Priority Flows 1 2 3 4 Low Priority Any static mapping mechanism degrades performance!
Transport Strategies in Unison In-network Prioritization + Arbitration Idea Arbitrator As a flow’s turn comes, map it Dynamic mapping of to the highest priority queue! flows to queues
Transport Strategies in Unison In-network Prioritization + Arbitration Idea Arbitrator As a flow’s turn comes, map it Dynamic mapping of to the highest priority queue! flows to queues High Priority High Priority Flows 1 2 3 Arbitrator 4 Low Priority Time t 1
Transport Strategies in Unison In-network Prioritization + Arbitration Idea Arbitrator As a flow’s turn comes, map it Dynamic mapping of to the highest priority queue! flows to queues High Priority High Priority High Priority High Priority Flows 1 Flows 2 2 3 3 Arbitrator Arbitrator 4 4 Low Priority Low Priority Time t 1 Time t 2
Transport Strategies in Unison In-network Prioritization + Arbitration Idea Arbitrator As a flow’s turn comes, map it Dynamic mapping of to the highest priority queue! flows to queues Similarly, High Priority High Priority • Arbitration + Self-Adjusting Endpoints • Arbitration + Self-Adjusting Endpoints High Priority High Priority Flows 1 Flows • Arbitration + In-network Prioritization 2 2 PASE leverages these insights in its design! 3 3 Arbitrator Arbitrator 4 4 Low Priority Low Priority Time t 1 Time t 2
Rest of the Talk … DC Transport Strategies PASE Design Evaluation
PASE Design Principle Each transport strategy should focus on what it is best at doing! Arbitrators – Do inter-flow prioritization at coarse time-scales Endpoints – Probe for any spare link capacity In-network prioritization – Do per-packet prioritization at sub-RTT timescales
PASE Overview Sender Receiver Arbitrator
PASE Overview Sender Receiver Arbitrator Arbitration: Control plane Calculate “reference rate” and “priority queue”
Recommend
More recommend