About The Forum
Calendar 2017 Poster Sessions : Homa: A Receiver-Driven Low-Latency Transport ProtocolUsing Network Priorities
Abstract:
Homa is a new datacenter transport protocol that provides
low latency for small messages while also supporting large
messages and high network utilization. In Homa, incoming
packets are scheduled by receivers in order to prioritize
shorter messages and manage congestion; senders transmit
small amounts of unscheduled data to avoid an extra round-
trip of latency for scheduling. Homa uses in-network
priority queues for efficient preemption and, more importantly,
to implement a form of bipartite matching between senders
and receivers, which ensures high network utilization.
Simulations show that Homa delivers short messages with
latency as low as pFabric and better than pHost under high
loads, and it can operate at a higher network utilization than
either pFabric or pHost.
Bio:
Behnam Montazeri is a PhD Candidate at Stanford Electrical Engineering department, working on HomaTransport and RAMCloud storage system under supervision of Prof. John K. Ousterhout in the PlatformLab in Stanford Computer Science department. His research spans the areas of distributed systems and low latency datacenter networks. Prior to RAMCloud, he worked on the Dielectric Laser Accelerator project as a joint member of the Byer group at Stanford's Ginzton laboratory and the Advanced Accelerator Research Department at SLAC National Accelerator Laboratory.
He holds a B.Sc. in Electrical Engineering from Isfahan University of Technology in Isfahan, Iran and an MS degree in Electrical Engineering from Stanford University.
Homa is a new datacenter transport protocol that provides
low latency for small messages while also supporting large
messages and high network utilization. In Homa, incoming
packets are scheduled by receivers in order to prioritize
shorter messages and manage congestion; senders transmit
small amounts of unscheduled data to avoid an extra round-
trip of latency for scheduling. Homa uses in-network
priority queues for efficient preemption and, more importantly,
to implement a form of bipartite matching between senders
and receivers, which ensures high network utilization.
Simulations show that Homa delivers short messages with
latency as low as pFabric and better than pHost under high
loads, and it can operate at a higher network utilization than
either pFabric or pHost.
Bio:
Behnam Montazeri is a PhD Candidate at Stanford Electrical Engineering department, working on HomaTransport and RAMCloud storage system under supervision of Prof. John K. Ousterhout in the PlatformLab in Stanford Computer Science department. His research spans the areas of distributed systems and low latency datacenter networks. Prior to RAMCloud, he worked on the Dielectric Laser Accelerator project as a joint member of the Byer group at Stanford's Ginzton laboratory and the Advanced Accelerator Research Department at SLAC National Accelerator Laboratory.
He holds a B.Sc. in Electrical Engineering from Isfahan University of Technology in Isfahan, Iran and an MS degree in Electrical Engineering from Stanford University.