Integrating Network Management For Cloud Computing Services
Cloud computing is known to lower costs of corporate IT. Thus enterprises are eager
to move IT applications into public or private cloud. Because of this trend, networks
connecting enterprises and cloud providers now play a critical role in delivering highquality
Simply buying better devices is not viable for improving network quality, due to
high capital costs. A more attractive approach is to better utilize network resources
with proper network management. However, there are two problems with current
network management: separately managing network components along the end-toend
path, and heavily relying on vendor-specific interfaces with devices.
This dissertation takes a practical approach driven by operational experiences of
cloud services to tackle the two problems. With knowledge of real-world challenges,
we have designed proper abstractions for low-level device interactions, and have built
efficient and scalable systems to integrate the management of various network components.
With commercial deployment, our operational experiences feed back into
revision of the abstraction and system design.
In this dissertation, we make three major contributions. We first propose to
consolidate the traffic and infrastructure management in datacenters. Our system,
called Statesman, simplifies management solutions by providing a uniform abstraction
to interact with various aspects of devices. Statesman then allows multiple solutions
to run together, resolves their conflicts, and prevents network-wide failures caused
by their collective actions. Statesman has been operational worldwide in Microsoft’s
public cloud offering since October 2013.
The second contribution consists of joining end hosts with networks for cooperative
traffic management. Our Hone system brings in the fine-grained knowledge of cloud
applications in the hosts, and offers an expressive programming framework with a
uniform view of both host and network data. Hone has been integrated into Verizon
The final contribution consists of bridging enterprises and Internet service
providers (ISPs) for fine-grained control of inbound traffic from cloud applications.
Our Sprite system enables enterprises to directly decide how traffic enters the enterprise
networks via which ISPs, offering expressive interface and scalable execution.
In collaboration with Princeton’s Office of Information Technology, Sprite was tested
with campus-network data and live Internet experiments.