Network science and engineering research has reached an inflection point with switch hardware vendors (merchant silicon) and network equipment manufacturers serving the software-defined networking (SDN) paradigm with a programmable forwarding plane. Measurement-feedback mechanisms must be developed to manage networks more effectively in an SDN environment utilizing the programmable flow definitions.

For example, load balancing between compute resources and protocol performance monitoring are now possible at runtime. Network performance may become deterministic for a workload on a set of compute resources when flow definitions govern the networking.

Dr. Deniz Gurkan’s network research provides an invaluable bridge between the SDN paradigm in networking and computations executed on HPC resources. SDN is based on the programmable data plane where control over path selection is centralized and forwarding elements (virtual and physical switches) have programmable flow tables.

As a result, a network operating system is emerging to help applications (such as compute jobs in bioinformatics, genomics, etc.) utilize networks efficiently while still parallelizing and optimizing processing speeds.

Traditionally, network optimization has been detached from compute needs with a socket API. Furthermore, merchant silicon for switch hardware has limited the innovation in handling application-specific flows for compute jobs.

Dr. Gurkan is advancing SDN to bridge forwarding (data) plane programmability with a domain-specific compute job’s network usage. Requirements from a compute application’s network usage will be load-balanced on both network and compute resources seamlessly with the centralized view of flow definitions.

This research will enable controller applications on HPC resources to monitor network performance and control flows at runtime so that overall performance improves for the compute jobs.

Her investigative research questions include: Hadoop processes have a map-reduce phase over a set of networked compute resources. How are these resources utilized in a load-balanced manner with optimal network efficiency? Is there a need for network optimization for all map-reduce phases?

HPC resources may run Dr. Gurkan’s network optimization applications transparently with compute jobs submitted by domain scientists while improvements in network performance are tested and verified to advance technologies within the SDN paradigm.