1: \begin{abstract}

2: The conventional approach to scaling Software-Defined Networking (SDN) 

3: controllers today is to partition switches based on network topology, with each 

4: partition being controlled by a single physical controller, running all SDN applications.

5: However, topological partitioning is limited by the fact that

6: (i) performance of latency-sensitive (e.g., monitoring) SDN applications 

7: associated with a given partition may be impacted by co-located 

8: compute-intensive (e.g., route computation) applications; (ii) simultaneously

9: achieving low convergence time and response times might be challenging;

10: and (iii) communication between instances of an application across partitions

11: may increase latencies. To tackle these issues, in this paper, 

12: we explore  {\em functional slicing}, a complementary approach to scaling,

13: where multiple SDN applications belonging to the same topological partition 

14: may be placed in physically distinct servers. We present \textit{Hydra}, a framework 

15: for distributed SDN controllers based on functional slicing. Hydra chooses

16: partitions based on convergence time as the primary metric, but places application

17: instances across partitions in a manner that keeps response times low while

18: considering communication between applications of a partition, and instances of

19: an application across partitions. Evaluations using the Floodlight controller

20: show the importance and effectiveness of Hydra in simultaneously keeping

21: convergence times on failures small, while sustaining higher throughput per

22: partition and ensuring responsiveness to latency sensitive applications.

23: \end{abstract}

24: