• Aucun résultat trouvé

User Association in Heterogeneous Wireless Networks (Chap- (Chap-ter 6)(Chap-ter 6)

In addition to using third-party WiFi or own WiFi platforms, operators are increasingly considering denser, heterogeneous network (HetNet) cellular deployments. In a Het-Net, a large number of small cells (SC) are deployed along with macrocells to improve spatial reuse [55, 56, 57]. The higher the deployment density, the better the chance that a user equipment (UE) can be associated with a nearby base station (BS) with high signal strength, and the more the options to balance the load. Additionally, the line between out-of-band WiFi “cells”, and in-band cellular small cells (femto, pico), is getting blurred, due to developments like LTE-unlicensed. The trend towards base station (BS) densification will continue in 5G systems, towardsUltra-Dense Networks (UDN)where: (i) many different small cells (SCs) are in range of most users; (ii) a small number of users will be active at each SC [58, 59, 6]. The resulting traffic vari-ability makes optimal user association a challenging problem in UDNs [58], as the goal here is not to simply choose between WiFi or cellular as in the previous chapter, but to choose between a large range of heterogeneous and overlapping cells.

As a result, a number of research works emerged that studied the problem of user association in heterogeneous networks, optimizing user rates [60, 61], balancing BS loads [62], or pursuing a weighted tradeoff of them [63]. Range-expansion techniques, where the SINR of lightly loaded BSs is biased to make them more attractive to the users are also popular [56, 57]. The main goal of these works is to offload or “steer”

traffic away from overloaded base stations, towards underloaded ones, while maintain-ing (or improvmaintain-ing) user performance. Nevertheless, the majority of these works fo-cused on DL traffic and the radio access link only. Future user association algorithms should be sophisticated enough to consider a number of other factors.

Uplink: While optimization of current networks revolves around the downlink (DL) performance, social networks, machine type communication (MTC), and other upload-intensive applications make uplink (UL) performance just as im-portant. Some SCs might see their UL resources congested, while others their DL resources, depending on the type of user(s) associated with that SC. What is more, the same SC might experience higher UL or DL traffic demand over time.

Traffic Classes: Most existing studies of user association considered homoge-neous traffic profiles. For example, [63, 64, 65] assume that all flows generated by a UE are “best-effort” (or “elastic”). Modern and future networks will have to deal with high traffic differentiation, with certain flows being able to require specific,dedicatedresources [66], [67, 68]. Such dedicated flows do not “share”

BS resources like best-effort ones, are sensitive to additional QoS metrics, and affect cell load differently.

Backhaul Network: Ignoring the backhaul during user association is reasonable for legacy cellular networks, given that the macrocell backhaul is often over-provisioned (e.g., fiber). However, the considerably higher number of small cells, and related Capital Expenditure (CAPEX) and Operational Expenditure (OPEX) suggest that backhaul links will mostly be inexpensive wired or wireless (in licensed or unlicensed bands), and underprovisioned [69]. Multiple BS might also have to share the capacity of a single backhaul link due to, e.g, point-to-multipoint (PMP) or multi-hop mesh topologies to the aggregation node(s) [70].

Hence, associating a user to a given BS might lead to backhaul congestion and low end-to-end performance, even if that BS can provide a high radio access rate to the user.

Chapter Contributions

In the following three works, we have addressed these exact questions, within a com-mon unifying framework [71, 72, 73].

N. Sapountzis, T. Spyropoulos, N. Nikaein, and U. Salim,“An Analytical Frame-work for Optimal Downlink-Uplink User Association in HetNets with Traffic Dif-ferentiation,”in Proc. of IEEE GLOBECOM 2015.

N. Sapountzis, T. Spyropoulos, N. Nikaein, U. Salim, “Optimal Downlink and Uplink User Association in Backhaul-limited HetNets,”in Proc. of IEEE INFO-COM 2016.

N. Sapountzis, T. Spyropoulos, N. Nikaein, U. Salim,User Association in Het-Nets: Impact of Traffic Differentiation and Backhaul Limitations,”in ACM/IEEE Transactions on Networking, 25(6): 3396-3410, 2017.

The related contributions can be summarized as follows:

Contribution 6.1Using the flow-level performance framework of α-fair user asso-ciation [63] as our starting point, we extended it considerbaly to include (i) traffic differentiation: splitting scheduler resources between GBR and non-GBR traffic, and assuming a different scheduler for each class; (ii) UL traffic: considering both legacy scenarios where UL and DL traffic must go through the same BS, as well as the envi-sioned UL/DL split [74]; (iii) backhaul characteristics: considering underprovienvi-sioned backhaul links as well as different network topologies (star or tree).

Contribution 6.2We derived optimal association rules for each combination of the above scenarios that can be implemented efficiently in a distributed manner and proved convergence to the global optimum (i.e., without any central element requiring global knowledge of all network state).

Contribution 6.3We used extensive simulations to understand the impact of backhaul topology, and interplay of UL/DL traffic, and showed that our algorithm outperformed state of the art algorithms at that time.

In addition to this work, presented in detail in Chapter 6, we have very recently made new contributions related to user association in heterogeneous, dense networks.

In the interest of space, we will not elaborate on them further in the respective chapter, but only provide a short summary here.

Joint Optimization of User Association and Flexible TDD .

Conventional networks usually operate with the same amount of resources for UL and DL (FDD or static TDD) [75]. Consider for example the standard FDD which uses some fixed, separate bands for uplink and for downlink. Our results from [73]

suggested that this can often be suboptimal. In TDD, it is possible to configure more resources for DL than UL (or vice versa, although not common), but each macro-cell is usually configured with the same ratio to be used permanently. However, recently proposedDynamic or Flexible TDDsystems can better accommodate UL/DL traffic asymmetry, by varying the percentage of LTE subframes used for DL and UL trans-mission [76]. Hence, more UL resources can be allocated to SCs with UL intensive users, and vice versa. Nevertheless, dynamic TDD introduces new challenges.

First,there is a strong interplay between user association and dynamic TDD poli-cies. Consider the simple example of Fig. 1.1, where a UL intensive user (1) and a DL intensive user (2) are both in range of a SC A (which is close) and a SC B (which is further away). Assume that each SC has initially the same amount of DL and UL resouces (50%). Both the DL and the UL user will connect to A, as it offers the best SINR. Now, notice that any change in the TDD schedule of SC A will hurt one of the two users. However, assume that A increases its UL resources to80%. There are now 8/5(i.e. 60%) more resources for the UL user, which could lead to60%higher rate.

Furthermore, assume that SC B increases its DL resources to80%. Connecting the DL user to B can increase the available resource blocks for her also by a factor of8/5. If the resulting SINR decrease has a smaller impact than this factor, thenboth users can win by revisiting both the TDD schedules and the association decisions of the two base stations, in coordination.

Figure 1.1: Interplay between user association and TDD configuration.

The previous example, while oversimplified, helps illustrate some of the dependen-cies at hand. One important omission in the above example, is that we ignored the DL-to-UL cross interference that might arise if nearby BSs have different schedules (see Fig. 1.1). E.g. a macro-cell transmitting on the DL, can really hurt a nearby SC transmitting on the UL [77]. Hence, excessive liberty in tuning UL and DL resources might hurt rather than help.

To this end, our main goal in a recent paper was to use the above framework to jointly optimize user association and TDD allocation per BS in order to:

Associate users with BSs to optimize a chosen user- or network- centric perfor-mance metric (e.g. spectral efficiency, load-balancing, etc.).

Choose the TDD UL/DL configuration for each SC to best match the UL/DL traffic demand for that metric.

Consider the TDD UL/DL configuration of nearby SCs to avoid cross-interference.

The respective paper is [78]

N. Sapountzis, T. Spyropoulos, N. Nikaein, and U. Salim,“Joint Optimization of User Association and Dynamic TDD for Ultra-Dense Networks,”in Proc. of IEEE INFOCOM 2018,

and the related contributions where the following:

Contribution 6.4We modified our analytical framework further, in order to include the possibility to tune the DL/UL schedule per BS, as well as capture cross-interference from nearby schedule discrepancies.

Contribution 6.5 We showed that the joint problem is non-convex in general, and propose a primal decomposition algorithm that reduces complexity and can be imple-mented in a distributed manner. We then prove that this algorithm converges to the optimal solution of the joint problem.

Contribution 6.6Using simulations, we show our approach canconcurrentlyimprove UL and DL performance compared to the state of the art, showing more than2× ag-gregate improvement, in the scenarios considered.