Michael Scharf

Network-aware service placement in a distributed cloud environment

We consider a system of compute and storage resources geographically distributed over a large number of locations connected via a wide-area network. By distributing the resources, latency to users can be decreased, bandwidth costs reduced and availablility increased. The challenge is to distribute services with varying characteristics among the data centers optimally. Some services are very latency sensitive, others need vast amounts of storage, and yet others are computationally complex but do not require hard deadlines on execution. We propose efficient algorithms for the placement of services to get the maximum benefit from a distributed cloud systems. The algorithms need input on the status of the network, compute resources and data resources, which are matched to application requirements. This demonstration shows how a network-aware cloud can combine all three resource types - computation, storage, and network connectivity - in distributed cloud environments. Our dynamic service placement algorithm monitors the network and data center resources in real-time. Our prototype uses the information gathered to place or migrate services to provide the best user experience for a service.

Abstracting network state in Software Defined Networks (SDN) for rendezvous services

The Software Defined Network (SDN) model depends on abstractions to separate the control plane from the packet forwarding plane. Applications can interact with the control plane to receive a global network view, upon which they can operate. By having access to network topology information, applications can optimize decisions related to service rendezvous, service fulfillment, service placement and service removal. The network is in the best position to provide guidance to a broad class of applications, including peer-to-peer systems, Content Distribution Network (CDN), and datacenter applications. In all these use cases proximity matters as peers need to rendezvous with other peers and users need to rendezvous with the best cloud application or best CDN server. We maintain that a solution for such a rendezvous problem should be an intrinsic component of the emerging SDN model. A specific instance of a protocol that abstracts network topology is the Application Layer Traffic Optimization (ALTO) protocol. ALTO provides applications an abstract view of the network and thus enables applications to leverage a network without exposing the network providers internal details or policies. We argue that ALTO provides a clean, mature, standards-based and powerful abstraction, which can be used by SDNs today to obtain network information for solving the rendezvous problem.

Performance comparison of scheduling algorithms for multipath transfer

Multipath transport protocols such as Multipath TCP can concurrently use several subflows to transmit a TCP flow over potentially different paths. Since more than one subflow is used, an efficient multipath scheduling algorithm is needed at the sender. The objective of the scheduler is to identify the subflow over which the current data packet should be sent. This paper compares the most important types of schedulers for multipath transfers. We model their performance analytically and derive key metrics, most notably the resulting end-to-end delay over heterogeneous paths. Our results show that a scheduler minimizing the packet delivery delay yields the best overall performance, but it is complex to realize. An alternative scheduler based on the sender queue size is simpler and has sufficient performance for relatively small asymmetry between the multiple paths. Our model results are confirmed by measurements with a real multipath transport protocol.

Comparison of end-to-end and network-supported fast startup congestion control schemes

A fundamental challenge for congestion control is the flow startup phase after the connection setup or after long idle periods. The Transmission Control Protocol (TCP) uses the time-consuming Slow-Start mechanism in that case. The objective of fast startup congestion control is to fully utilize a path much more rapidly. Faster startups can either be realized by new end-to-end congestion control mechanisms that change the Slow-Start algorithm, or by additional on-path signaling providing information about the path. This paper comprehensively compares end-to-end and network-supported fast startup schemes. We evaluate both the potential performance benefit and risk, using analytical models, simulation results, and experiments with real applications. Concerning end-to-end schemes, we study both a simple increase of the initial congestion window and the recently proposed Jump-Start mechanism. The considered network-supported schemes are the Quick-Start TCP extension and the Rate Control Protocol (RCP). They can all significantly reduce the transport delay of mid-sized data transfers, in particular over broadband networks with a non-negligible latency. As a result, fast startup schemes can significantly speed up Internet application with such traffic patterns. Our results also reveal that end-to-end approaches are not necessarily overly aggressive. Network-supported mechanisms can reduce the risk of congestion, but there are several unsolved open issues. In summary, we argue that end-to-end fast startup schemes such as an increased initial congestion window would be a promising solution for the Internet.

Monitoring and abstraction for networked clouds

We consider the problem of building tightly coupled network and cloud management systems for “carrier clouds” based on an abstract view of the dynamic network state. Optimized resource placement in distributed clouds requires information about the internal network topology and state, in addition to other data center information. We present a distributed cloud system that accesses such data in an abstract way using the Application Layer Traffic Optimization (ALTO) protocol. The demonstration addresses how dynamic networking information can be gathered, how it can be abstracted and coupled with other data such as data center load, how it can be exposed, and how it can be integrated into a cloud management system. Our distributed cloud solution demonstrates that ALTO is well-suited for infrastructure-to-application information exposure.

MCTCP: A Multipath Transport Shim Layer

Multipath transport protocols enable the concurrent use of different paths and resource pooling of their capacity. Ongoing research and standardization activities extend the Transmission Control Protocol (TCP) towards a Multipath TCP solution. The objective is to aggregate multiple subflows between two endpoints into a single session that can be accessed by an application like a single TCP connection. This paper presents and evaluates Multi-Connection TCP (MCTCP) as a new multipath transport solution. MCTCP consists of a shim layer on top of several TCP connections and encodes control information, as far as possible, in their payload. In this paper, we first motivate MCTCPs design and compare it to a multipath transport protocol that uses TCP option encoding only. Second, we show that MCTCP can be implemented in the Linux stack with very few kernel modifications. We also demonstrate that MCTCP can be combined with a coupled multipath congestion control mechanism. Measurement results prove that MCTCP is an efficient and robust protocol that can dynamically pool the bandwidth of several paths.

Dynamic VPN Optimization by ALTO Guidance

Virtual Private Networks (VPNs) are a key component of cloud computing systems, since they provide isolated connectivity between geographically separated users. The elasticity in cloud computing and new usage patterns such as cloud bursting require VPNs to be more dynamic than traditional solutions used by network service providers. Managing and optimizing the topology of VPNs requires insight into the underlying wide area network topology and benefits from new network interfaces currently discussed for Software Defined Networks (SDN). This paper presents the use of the Application-Layer Traffic Optimization (ALTO) protocol for VPN optimization. ALTO is a standardized solution for exposure of abstract topology information to a variety of applications, including cloud management systems. We demonstrate the use of ALTO in determining how to scale-out a VPN on demand. We also present a prototype of an ALTO-based dynamic VPN management, which is based on a carrier-grade network management system. Our results show that ALTO is a powerful topology abstraction approach that enables informed VPN scale-out decisions by applications.

Network-Aware Instance Scheduling in OpenStack

Cloud computing systems require a placement logic that decides where to allocate resources. In state-of-the-art platforms such as OpenStack, this scheduler takes into account multiple constraints when starting a new instance, including in particular the required computational and memory resources. However, this scheduling mechanism typically neither considers network requirements of Virtual Machines nor the networking resources that are actually available. In this paper we present an extension of the OpenStack scheduler that enables a network-aware placement of instances by taking into account bandwidth constraints to and from nodes. Our solution keeps track of host-local network resource allocation, and it can be combined with bandwidth enforcement mechanisms such as rate limiting. We present a prototype that requires only very few changes in the OpenStack open source software. Testbed measurement results demonstrate the benefit of our solution compared to the OpenStack default approach.

ATLAS: Accurate Topology Level-of-Detail Abstraction System

The ability to extract topology information from the network is important for many applications and enables more informed resource selection. The challenge for topology exposure is to provide a compact representation that is sufficiently accurate and complies to topology hiding policies. This paper presents a topology abstraction system that can expose large-scale service provider network maps with an adjustable level-of-detail. Our system uses graph sparsification algorithms to reduce the complexity of routing topologies. Our numerical results reveal that the size of maps can be reduced by one order of magnitude or more while the result still enables reasonable traffic optimization inside applications. A proof-of-concept implementation gathers network management system data and exposes abstract maps through the Application-Layer Traffic Optimization (ALTO) protocol.

Sparsifying network topologies for application guidance

Topology managers expose network information to applications to improve application-level resource management. An example for various ongoing standardization activities is Application-Layer Traffic Optimization (ALTO). Due to privacy and security constraints, exposed network information has to be filtered according to policies. As part of such policies, distance information can be abstracted by presenting coarser distances over pairs of clustered nodes rather than the precise distances between all node pairs. We refer to this process as distance sparsification. The contribution of this paper is two-fold, the first being a new policy system to enforce abstraction. The second and the main contribution addresses the algorithmic challenge. For the latter, we consider two types of distance sparsification algorithms. The first variant takes as input a matrix of pairwise distances. The sparsification algorithm produces a smaller distance matrix by clustering the nodes into clusters. The second variant instead collapses an edge-weighted graph. We measure the performance of the algorithms by the accuracy of the resulting sparsified distances, and we show that matrix sparsification outperforms graph sparsification. We further observe the trade-off between the accuracy and the size of the sparsified representation. In addition, we also extend our algorithms to handle labeled data, i. e., abstraction policies explicitly mark a number of destinations as reference points. Such additional information can improve the distance sparsification.