[talks] M Yu preFPO
mml at CS.Princeton.EDU
Tue Nov 16 11:23:28 EST 2010
Minlan Yu will present her preFPO on Wednesday November 24 at 9AM in Room 402.
The members of her committee are: Jennifer Rexford, advisor; Mike Freedman and
Albert Greenberg (Microsoft), readers; Vivek Pai and Ed Felten, nonreaders. Everyone is
invited to attend her talk. Her abstract follows below.
Title: Supporting Flexible Policies in Large Enterprise and Data-Center Networks
Enterprise and data center networks are growing larger and becoming
more complicated to manage. These networks consist of two layers: the
management layer that defines routing, enforces high-level policies, and
diagnoses performance problems; the data plane that forwards packets,
applies fine-grained rules, and monitors traffic.
To make management layer simpler and easier to operate, people have
proposed flat addressing to
support host mobility, shortest-path routing to improve performance,
and flow-based switches to enable fine-grained policies. However, these
techniques raise scalability challenges in the data plane with many
switches, flow rules, and applications, forcing the use of expensive
switches with large, power-hungry memory for processing and monitoring
This thesis proposes a new scalable data plane for enterprises and data centers.
We improve the scalability of enterprise networks through a
combination of new data structures that make effective use of limited
memory, traffic indirection that minimizes the state required at each
switch, and end-host based monitoring to reduce the overhead at switches.
Our proposed system consists of three parts:
(i) BUFFALO: A scalable packet forwarding architecture that uses Bloom
filters to reduce the switch memory usage, and gracefully handles
false positives by sending packets through a slightly longer path.
(ii) DIFANE: A distributed management architecture that
leverages emerging flow-based switches to
scalably forward packets based on policies.
(iii) SNAP: A network performance diagnosis architecture for data
center applications that passively logs traffic statistics in the
end-host network stack
and pinpoint problems that occur at the network device, network stack
and the application software.
BUFFALO and DIFANE can be easily implemented with small modifications in today's
switches, as demonstrated by our prototypes built using the Click
modular router. SNAP was deployed in a large data center networks and
helped operators and developers to pinpoint performance problems.
More information about the talks