Congestion integrated control in virtualized clouds

Abstract

Many cloud computing platforms adopt visualization technology to provide cloud services. To achieve high computing performance, many data parallel frameworks are adopted to run the jobs. The job is broken into pieces and farmed out to virtual machines (VMs) to process in parallel. The responses from these VMs are aggregated to produce the final result with concurrent flows. However, the concurrent flows adjust congestion window in an isolated manner, there will be large variance among the transmission time of each flow. It will have negative impact on the cloud application performance since the slowest flow determines the total completion time of concurrent flows. To facilitate resource utilization and VM management, these VMs usually reside in the same physical host which offers the native infrastructure for concurrent flows' integrated management. Therefore, we propose a Congestion Integrated Control (CIC) scheme which offloads each flow's independent congestion control module and sets up an integrated control platform to perform the congestion control for concurrent flows in a unified manner. The CIC scheme is implemented on the Xen platform. The performance evaluation indicates that the large variance among transmission time of each flow is substantially reduced and total completion time of concurrent flows is decreased correspondingly.