Virtualization is central to cloud computing systems. It abstracts computing resources to be shared among multiple virtual machines (VMs) that can be easily managed to run multiple applications and services. To benefit from the advantages of cloud computing, and to cope with increasing traffic demands, telecom operators have adopted cloud computing. Telecom services and applications are, however, characterized by real-time responsiveness, strict end-to-end latency, and high reliability. Due to the inherent overhead of virtualization, the network performance of applications and services can be degraded. To improve the performance of emerging applications and services that demand stringent end-to-end latency, and to understand the network performance bottleneck of virtualization, a comprehensive performance measurement and analysis is required. To this end, we conducted controlled and detailed experiments to understand the impact of virtualization on end-to-end latency and the performance of transport protocols in a virtualized environment. We also provide a packet delay breakdown in the virtualization layer which helps in the optimization of hypervisor components. Our experimental results indicate that the end-to-end latency and packet delay in the virtualization layer are increased with co-located VMs.

Network Function Virtualization (NFV) is a promising solution for telecom operators and service providers to improve business agility, by enabling a fast deployment of new services, and by making it possible for them to cope with the increasing traffic volume and service demand. NFV enables virtualization of network functions that can be deployed as virtual machines on general purpose server hardware in cloud environments, effectively reducing deployment and operational costs. To benefit from the advantages of NFV, virtual network functions (VNFs) need to be provisioned with sufficient resources and perform without impacting network quality of service (QoS). To this end, this paper proposes a model for VNFs placement and provisioning optimization while guaranteeing the latency requirements of the service chains. Our goal is to optimize resource utilization in order to reduce cost satisfying the QoS such as end- to-end latency. We extend a related VNFs placement optimization with a fine-grained latency model including virtualization overhead. The model is evaluated with a simulated network and it provides placement solutions ensuring the required QoS guarantees. 

Datacenter applications generate a mix of short and long flows, which have often contrasting network performance requirements. While short flows are typically sensitive to their completion time, long flows are more or less deadline agnostic but demand high throughput. Despite the availability of multiple, parallel high-capacity paths inside a datacenter network, the achievable transport-layer performance for both latency-sensitive and capacity-demanding applications is far from optimal. The reason is partly due to the inefficiency of transport protocols deployed inside datacenters. Existing transport protocols are either not capable of utilizing multiple paths offered by datacenter topologies, e.g., Datacenter TCP (DCTCP) or unsuitable for latency-sensitive applications, e.g., Multipath TCP (MPTCP), due to the employed congestion detection schemes. To address this problem, we have designed a coupled multipath congestion control algorithm called Multipath Datacenter TCP (MDTCP). MDTCP builds upon MPTCP and uses Explicit Congestion Notification (ECN) signals to detect and react to congestion before queues overflow as in DCTCP, offering both reduced latency and higher network utilization. The MDTCP congestion control has been implemented in the Linux kernel and in a packet- level network simulator. We evaluate MDTCP’s performance extensively both in a programmable datacenter network testbed andin large-scale simulations. The obtained results show that MDTCP outperforms DCTCP by reducing the average Flow Completion Time (FCT) by more than 1.6× at high load, and achieves similar performance as DCTCP at moderate network load. Moreover, it outperforms MPTCP by always achieving a lower average FCT. MDTCP also improves network utilization by 7% and 12% compared to MPTCP and DCTCP, respectively.

The strict low-latency requirements of applications such as virtual reality, online gaming, etc., can not be satisfied by the current internet. This is due to the characteristics of classic TCP such as Reno and TCP Cubic which induce high queuing delays when used for capacity-seeking traffic, which in turn results in unpredictable latency. The Low Latency, Low Loss, Scalable throughput (L4S) architecture addresses this problem by combining scalable congestion controls such as DCTCP and TCP Prague with early congestion signaling from the network. It defines a Dual Queue Coupled (DQC) AQM that isolates low-latency traffic from the queuing delay of classic traffic while ensuring the safe co-existence of scalable and classic flows on the global Internet. In this paper, we benchmarktheDualPI2 scheduler, a reference implementation of DQC AQM, to validate some of the experimental result(s) reported in the previous works that demonstrate the co-existence of scalable and classic congestion controls and its low-latency service. Our results validate the co-existence of scalable and classic flows using DualPI2 Singlequeue (SingleQ) AQM, and queue latency isolation of scalable flows using DualPI2 Dual queue (DualQ) AQM. However, the rate or win-dow fairness between DCTCP without fair-queuing (FQ) pacing and TCP Cubic using DualPI2 DualQ AQM deviates from the original results. We attribute the difference in our results and the original results to the sensitivity of the L4S architecture to traffic bursts and the burst sending pattern of the Linux kernel.

Queuing latency is one of the limiting factors to achieve the latency targets of emerging latency-sensitive Internet applications (e.g., interactive web, real-time online gaming). It occurs when large capacity-seeking traffic bloats router buffers configured to allow full link utilization of standard TCP congestion controllers (e.g., TCP Reno, Cubic). The Low Latency, Low Loss and Scalable Throughput (L4S) architecture proposes to overcome the problem by combining scalable congestion controllers (e.g., DCTCP, TCP Prague) and early congestion signaling from the network. L4S defines Dual Queue Coupled (DQC) AQM as transition mechanism enabling scalable senders to coexist with standard congestion control. This paper extends the L4S Internet service to the multipath domain by using MDTCP—a scalable multipath congestion control for Multipath TCP (MPTCP). We evaluate the performance of MDTCP in a controlled network environment mimicking the L4S Internet service architecture. Our results indicate that MDTCP and TCP Prague achieve similar flow completion time for short flows, and outperform both the non-scalable, single-path, and multipath congestion controls. MDTCP also improves multipath capacity utilization compared to the existing MPTCP congestion controllers and outperforms both TCP Prague and TCP Cubic. Although MDTCP achieves a lower FCT for medium flows than TCP Prague, it does not show improved performance than the classic CCs due to its exit from the slow-start phase. With regard to bottleneck capacity sharing, MDTCP never caused starvation when sharing a bottleneck with a single-path TCP Prague, and is not severely affected by the competing TCP Prague flows.