Dharmesh Kakadia and Vasudeva Varma, SIEL, IIIT Hyderabad [email protected] IBM Collabora9ve Academia Research Exchange (I-‐CARE), Oct’ 12 Parameter Value Parameter Value Number of Hosts 2000 Number of Ports per Switch 24 Switch booEng Eme 90 sec Number of Edge Switches 100 Topology FatTree Link Capacity 100 MBPS Motivation • Large operational cost (OPEX) of infrastructure • Network power contributes 15% of the amortized cost • Turning off a port does not help much: power consumption of a switch varies less than 8% when utilization varies from zero to full • SLAs are important Conclusion • Near linear power consumption of network • Scalable • Stable Ongoing and Future Work • Implementation on OpenStack • Validation on different topologies and traffic • Quantifying the effect on fault-tolerance • Other approaches to prioritise flows Software Defined Networking • Separation of control and data plane functionalities • Control plane is implemented in software Proposed Algorithm OptimizeAllocation(S){ Update traffic stats using SDN counters For each Switch s in S such that Utilization(s) < threshold Θ over time t do { if(canMigrate(s, S-s)){ pFlows=prioritiseFlows(s) incrementalMigration(pFlows) Power-off(s) } } } Data center Networks • Designed for peak load and always on assumption • Does not consider traffic variation - < 25% links are hotspots - Huge traffic variation during different time of day. - 75% of traffic stays within a rack • Effective routing algorithm to reduce utilization • Load balance across paths and migrate VMs Results Experimental setup • Simulation using Mininet • Floodlight as SDN controller • Random traffic from each host to fixed no of other hosts • Delay variation as an indicator of SLA adherence OpenFlow • Open networking interface – one way of achieving SDN • Centralized view of network • Remotely controlling forwarding tables of network devices.