THESIS
2016
xxii, 124 pages : illustrations ; 30 cm
Abstract
The rapid growth of cloud computing, network virtualization and big data brings new
challenges for computer networks. By decoupling the control plane and the data plane,
software defined networking becomes an emerging paradigm to enable network innovation
with unprecedented programmability. The major concerns are performance issue for large
networks and how to facilitate network management by SDN visibility.
Initially, we explore the extra flow setup latency by the controller and switch communication.
To eliminate the overhead, we propose a system which predicts frequent
communication pair and proactively installs forwarding wildcard rules.
Then, we concentrate on software defined measurement and propose three novel schemes
to optimize network monitoring efficiency in a...[
Read more ]
The rapid growth of cloud computing, network virtualization and big data brings new
challenges for computer networks. By decoupling the control plane and the data plane,
software defined networking becomes an emerging paradigm to enable network innovation
with unprecedented programmability. The major concerns are performance issue for large
networks and how to facilitate network management by SDN visibility.
Initially, we explore the extra flow setup latency by the controller and switch communication.
To eliminate the overhead, we propose a system which predicts frequent
communication pair and proactively installs forwarding wildcard rules.
Then, we concentrate on software defined measurement and propose three novel schemes
to optimize network monitoring efficiency in a top-down approach. First, we present
a cross-layer optimization for sketch-based measurement. We observe the diminishing
marginal utility property of sketch-based measurement. By trading a little accuracy, we
dramatically shrink the measurement resource usage, and develop a two-stage heuristic
to efficiently assign concurrent measurement tasks to underlying switches.
Second, we propose schemes to optimize flow statistics collection. We point out flow
statistics polling is a fundamental primitive for software defined measurement. Based on this observation, we propose a generic optimization which is compatible with all existing
measurement frameworks. Two monitoring schemes are presented to achieve different
levels of measurement granularity.
Finally, we propose a measurement-aware controller placement which reduces the overhead
in the physical layer. Our proposal is cost-effective and application-agnostic. The
placement model takes both the synchronization and flow statistics polling cost into account.
Two heuristics are presented to efficiently generate near-optimal placements for
large-sized networks.
We demonstrate the effectiveness of our proposals by conducting experiments on various
network topologies with real-world traffic traces.
Post a Comment