Summary

As an offspring of ARPANET, the Internet has been designed as a packet network with robustness and survivability as first objectives. At that time, packet switching has been preferred over circuit switching. To cope with congestion due to an increasing popularity and the scarcity of resources, the Internet has then evolved with a fair-sharing objective implemented using the end-to-end principle. The global allocation performed by this system can be interpreted (cf. NUM) as a distributed optimization procedure aiming to maximize the individual utilities summed in an α-fair manner—α depends on the TCP variant.

Concurrently, the usage—video-conferencing, video-on-demand, large-file sharing, remote backup, data-intensive applications, e-science—and expectations of users regarding the Internet have also evolved, making both users and providers unhappy: the former because of the quality of service obtained, and the latter because their profit margins are impacted. Indeed, their revenues are not directly linked to the provided resources.

Variety of requirements and associated utility functions together with renewal of circuit through optical circuit switching and development of the automation of circuit provisioning in transport networks with unified control plane leads to consider other allocations mechanisms. They are suited for these new utilities such as variable provisioning of access link over time or deadline-constrained transfers and combination of them.

The most promising is based on services where users can express their needs through request submission. Then, by taking advantage of the packet switching paradigm and its relative fluidity, and providing the advantage of knowing in advance if the resource will be sufficient for the users, the service optimizes resource allocation under the constraints of both users and network operators. The existence of this service does not preclude classical best-effort approaches for the remainder of the traffic.

As an example, if a user wants a dedicated circuit for high-quality video conferencing, the request will contain different constraints specifying the requested QoS, such as a low delay—eg. less than 50 ms—and a constant bandwidth— Mbit/s for 4096*2048 px at a frame rate of 30.0 pict/s—for the duration of the video conference. Alternatively, users or data-intensive applications might want to have a deterministic completion of a 40 GB transfer within a time window of 10 min, and get an allocation fulfilling the constraint. This allocation is a profile of bandwidth that, over time, delivers the requested volume. The rest of the time, the needed bandwidth is much less.

In short, the problem is for operators to have an interest in providing users with the expected QoS. My contributions are distributed in five different sub-problems:

1. the evidence of the problem for bulk data transfers with the evaluation of transfer time predictability;
2. the algorithmics for the service that schedules requests of network utilization;
3. practical considerations on how to implement this service from the source application to the destination application, and its performance;
4. how a service provider can use a rented infrastructure to provide a transfer service;
5. finally, is this approach justified, from the points of view of users and service providers?

Predictability of Transfer Completion Time

Due to the dependence of the allocation to the number of participants, fair sharing shows a fundamental problem when it comes to predicting the allocation that one will get in capacity-constrained networks—although, even if the number of concurrent flows remains constant, it is still difficult. Using real experiments with different transport protocols, our first contribution shows that completion time under congestion depends on so many parameters that it is difficult—or impossible—for users to determine the completion time of their transfers .

Explicit Request and Scheduling of Transfers

To solve this, we propose a network-resource allocation service that allocates network bandwidth based on the users’ utilities while taking care of the value of the network. The allocation mechanism proposed gives both satisfaction to users—seeking best-effort or guaranteed services—and flexibility to network operators, which can plan in advance the resource utilization of users with large needs . The scheduler has been implemented in jBDTS .

Data Plane and Conformance to Allocated Profile

After the presentation of the bulk data transfer scheduling service on a static network, we present how bandwidth delivery can be articulated with the data plane to enable regular applications to efficiently use the allocated resources. For this purpose, we present the FLOC software and evaluate its performance regarding the original objective: being able to use this to enforce profiles of rates for flows .

Dynamic Bandwidth Provisioning and Bit Mover Service

Next, to benefit from dynamically reconfigurable networks, we introduce a tiered architecture: the clients or customers, the service provider, and the network provider. In this architecture, clients issue transfer requests to a bit-mover service that provisions the circuits of its dynamic infrastructure by issuing bandwidth requests to a network provider. We present the provisioning optimization problem as a linear program and simulations to highlight the benefit of mixing malleable requests with more constrained requests .

Extension of Routing Games to Flow over Time

Finally, we propose an extension of routing games with convex cost functions to consider time, as a tool to quantify, study and propose solutions to overcome the inefficiency introduced by selfishness in Wardrop-like games. As a result, the service provider is shown to be able to improve the social cost of the allocation, compared to the social cost of the Wardrop equilibrium reached by selfish decisions. This makes the service provider both legitimate and sustainable .

Download

PDF

This work is licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported License. Powered by Hakyll.