On Optimizing Data Transmission Latency In Datacenter Networks

Datacenters are large facilities that house tens of thousands of interconnected computing and storage servers. In recent years, large amount of investments have been made in datacenters to support services such as search, social networking and cloud computing. This has spurred a lot of interest in the industry and the research community in innovation for improving the performance of datacenter networks.Missions that run in private datacenter deployed with TCP needs long time to complete due to frequent flow timeout, which is caused by hourly traffic burst. Base on the simulation and logs of traffic burst, we find that the length of link idle period caused by timeout determines mission completion time. Combined with the analysis of TCP fast retransmission algorithm, we indicate that unfair window evolution of senders gives rise to long idle period. Then we provide a distributed rate control algorithm named MIP(Minimizing Idle Period) to improve the fairness of window increasement. The simulation results show that MIP obtains almost optimal mission completion time even under highly burst.Network delay is required to be guaranteed in multi-tenant cloud datacenters. In-stead of running TCP, cloud network provides minimum bandwidth guarantee so that the tenants’applications achieve predictable latency. There are two parts of work in bandwidth guarantee, on one hand, the virtual network model must represent the prop-erties of traffic in tenant’s application. On the other hand, physical links in data center networks should reserve as less bandwidth as possible when mapping the tenant re-quest into datacenter servers. However, recent researches do not consider the impact of middlebox on network latency. We specify the abstraction of tenant requirement as VMBA(Virtual MiddleBox Application) model and design a virtual machine place-ment algorithm for both middlebox and application VMs, which is called MBCloud. Extensive simulation results show that MBCloud could handle 11.3% more tenant than the baseline under same VM utilization.The main contribution of this paper includes three parts:(1)we summarize the reasons of long mission completion time; (2)we propose an algorithm for private data center that reduces mission completion time; (3)we design and implement a middle-box and VM placement algorithm that provide guaranteed delay for cloud data center. Our algorithms insure the performance in term of latency through the verification of simulation results.