Abstract
Minimizing the flow completion time (FCT) is widely considered as an important optimization goal in designing data center networks. However, existing schemes either rely on the precondition that the size and deadline of each flow is known in advance, or require modifying the switch hardware, which is hard to implement in practice. In this paper, we present MCPF, a flexible and dynamic flow scheduling strategy to reduce the FCT. This strategy is based on the estimated probabilities of each flow to finish the transmission in a period time, and these flows which have higher completion probabilities are assigned with higher priority. Meanwhile, switches perform flow scheduling according to these priorities. We employ a queueing theory based mathematical model to analyze the average FCT of MCPF, and compare it with other two flow scheduling strategies. We also introduce the challenges and the solutions to implement MCPF in realistic networks. Finally, we evaluate the performance of MCPF in Mininet. The analysis and experimental results show that MCPF could effectively reduce the FCT.
Highlights
IntroductionE.g. web search, social networks, retail, have become an indispensable part of human society
Online services, e.g. web search, social networks, retail, have become an indispensable part of human society
The primary conception of MCPF is flow completion probability, which reflects the possibility that a flow finish its transmission in a period time, and the order of flow scheduling depends on these probabilities of each flow
Summary
E.g. web search, social networks, retail, have become an indispensable part of human society. Mice flows are generally from interactive foreground services, and they are sensitive to the network latency These flows should be finished in a short time, while elephant flows don’t need to satisfy the deadline. It’s worthy of noting that MCPF strategy is an approach that can decrease FCT through flexible and dynamic flow scheduling in switches, and it is not incompatible with other types of schemes, such as DCTCP [3] and MPTCP [10]. It can further improve the performance of data center networks by synthesizing with other schemes
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have