Dcnet: Scalable Data Center
Abstract
Enterprise computing has moved to the cloud steadily over the past decade, and has imposed new requirements that mandate fundamental changes in networking infrastructure. Observation of data center traffic patterns shows that as applications move to virtual machines (VM) and containers, a significant amount of data center network communication occurs among servers within the data center, increasing the so called east-west traffic. Future data centers must scale beyond fifty and one-hundred Ggabits per second to Terabit per second network speeds. In addition to scalability and easy manageability, future data centers must also support efficient live VM migration. Migration of VMs across physical servers allows cloud providers to sup-port high compute loads, while improving power efficiency by eliminating hot spots, balancing loads, and performing preemptive preventative maintenance. We propose a data center net-work architecture that employs a revolutionary new approach, and supports live VM migration throughout an organization, including migration across an organization that spans multiple data centers. Our envisioned network system takes a radical departure by eliminating layer 2 functional-ity and reducing the protocol stack to achieve higher throughput and increased flexibility. Our approach works well in a data center because a data center offers a constrained environment managed by a single organization, which allows us to control packet transport completely. In place of layer 2 (MAC) addresses, the proposed DCnet paradigm defines a hierarchical tagging scheme, and arranges for smart NICs to transform the destination address in each outgoing or incoming datagram, mapping between a completely normal IP destination address and an equivalent routable tag used within the data center. Removing layer 2 from the network stack decreases overhead, eliminates encapsulation/de-encapsulation, simplifies buffer management, scales packet sizes for high-throughput networks, and thereby increases overall scalability.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 29, 2019
- Source ID
- HR00111910004
Entities
People
- Douglas Comer
Organizations
- Defense Advanced Research Projects Agency
- Purdue University