MPTCP under Virtual Machine Scheduling Impact

Abstract

Multipath TCP (MPTCP) has captured the networking community's attention in recent years since it simultaneously transfers data over multiple network interfaces, thus increases the performance and stability. Existing works on MPTCP study its performance only in traditional wired and wireless networks. Meanwhile, cloud computing has been growing rapidly with lots of applications deployed in private and public clouds, where virtual machine (VM) scheduling techniques are often adopted to share physical CPUs among VMs. This motivates us to study MPTCP's performance under VM scheduling impact. For the first time, we show that VM scheduling negatively impacts all MPTCP subflows' throughput. Specifically, VM scheduling causes the inaccuracy in computing the overall aggressiveness parameter of MPTCP congestion control, which leads to the slow increment of the congestion windows of all MPTCP subflows instead of just a single subflow. This finally results in a poor overall performance of MPTCP in cloud networks. We propose a modified version for MPTCP, which considers VM scheduling noises when MPTCP computes its overall aggressiveness parameter and its congestion windows. Experimental results show that our modified MPTCP performs considerably better (with up to 80% throughput improvement) than the original MPTCP in cloud networks.