Résumé:
Recent study showed that many network
applications require multiple different network flows
to complete their tasks. Provisioning bandwidth to
network applications other than individual flows in
data center networks is becoming increasingly important
to achieve user satisfaction on their received network
services. Modern data center networks commonly adopt
multi-rooted tree topologies. Equal-Cost Multi-Path
(ECMP) forwarding is often used to achieve high link
utilization and improve network throughput. Meanwhile,
max-min fairness is widely used to allocate network
bandwidth fairly among individual network flows.
Today’s data centers usually host diverse applications,
which have various priorities (e.g., mission critical
applications) and service level agreements (e.g., high
throughput). It is unclear how to adopt ECMP
forwarding and max-min fairness in the presence of
such requirements. In this paper, we first propose
a flow-based scheduling mechanism (called FlowSch)
to provide a prioritized Max-Min fair multiple path
forwarding to improve link utilization and improve
application performance. Then, we demonstrate and
discuss that FlowSch may not perform effectively when
network applications commonly use multiple network
flows to accomplish their tasks. Accordingly, we design
an application-aware scheduling mechanism (called
AppSch) to tackle this challenge. AppSch can optimally
allocate available bandwidth to satisfy application
requirements. Our performance evaluation results show
that FlowSch can improve flow throughput 10-12% on
average and increase overall link utilization especially
when the total demanded bandwidth is close or even
exceeds the bisectional bandwidth of a data center
network. However, when most applications rely on
multiple network flows, AppSch can improve link
utilization more effectively and reduce the application
completion time 36-58%.
