Performance Isolation for 5G RAN Slices Across Multiple Interfering Cells

Radio Access Network (RAN) refers to the wireless last-mile of cellular networks, connecting the base-station to the end-devices (referred as user equipments or UEs). We explain some relevant RAN concepts below.

As mentioned in §1, RAN slicing is a key 5G feature that brings the notion of multi-tenancy to cellular networks. With RAN slicing, the network operator (e.g. AT&T, Verizon, etc) divides the radio resources among heterogeneous slices (MVNOs, enterprises, applications, etc) as per their SLAs [25]. The SLA of a slice translates to a minimum quota of RBGs that the network operator must allocate to it at a base-station [12, 34]. Each slice then schedules its quota of radio resources across its own users as per its own performance objective (e.g. fairness, prioritization, throughput maximization, etc). The performance objective typically varies across slices [18, 21, 46]. A slicing system must support such diverse slice objectives, and ensure each slice gets its quota of resources while retaining performance isolation across slices.

Prior work on RAN slicing has largely been restricted to scheduling radio resources at a single base-station (e.g. [24, 34, 30, 12]). Other works [40, 10] look at the problem of mapping network-wide service-level agreements (SLAs) to per-cell resource quotas for a given slice but ignore potential interference across cells. We leverage these prior works for determining per-cell slice quotas and doing channel-aware slicing and scheduling [12] (as detailed in §3.1). The primary focus of our work, however, is to address the challenges that arise from the interplay of RAN slicing with interference management across multiple cells.

In terms of its goal, RAN slicing bears similarity to the classical concept of hierarchical fair scheduling, where we want weighted fairness across slices, and fairness (or other objectives such as prioritization or throughput maximization) within each slice. However, classical hierarchical scheduling assumes that the top-level allocator can reason about resource allocation independently of how those resources are used internally. In a wireless RAN, this assumption is already challenged even in a single-cell setting because user performance depends not only on how many radio resources (i.e., number of RBGs) are allocated, but also on the quality of the wireless channel associated with those RBGs. Moreover, channel quality varies across users for a given RBG, and across RBGs for a given user. This channel quality variation must be taken into account when assigning RBGs to different slices (and to different users within a slice). As a result, achieving hierarchical objectives efficiently requires channel-aware assignment of RBGs across slices (and across users within each slice), rather than treating RBGs as interchangeable units. §2.5 exemplifies how such channel-aware RAN slicing helps improve overall system throughput.

Our prior work, RadioSaber [12] proposed a system for channel-aware scheduling of radio resources at a given base-station across slices, and across users within each slice. However, RadioSaber makes a crucial simplifying assumption: the utility of a radio resource is local to a cell and does not depend on scheduling decisions made at other cells. This locality assumption holds in single-cell settings, but is violated in multi-cell deployments where wireless interference impacts scheduling decisions across cells.

In this work we focus on the effects of this wireless interference in multi-cell settings, that introduces new and unique challenges to the problem of hierarchical scheduling across slices. With interference, the quality associated with an RBG $R_i$ and cell $C_j$ depends not only on which slice (and which specific user within the slice) is scheduled on $R_i$ at $C_j$, but also on whether any user (belonging to any slice) is scheduled at $R_i$ at neighboring cells. In other words, the effective value of a radio resource is no longer local to a single cell. Because of this cross-cell scheduling dependency, we cannot perform channel-aware RAN slicing for each cell in isolation as was done by RadioSaber [12]. Instead interference management decisions must be made jointly with channel-aware scheduling for each RBG across all cells and across all slices. This coupling therefore implies that interference management decisions cannot be cleanly layered above slicing: whether muting a resource is beneficial or not depends on slice-specific objectives and intra-slice scheduling decisions. Consequently, this cross-cell dependency can easily undermine performance isolation across slices – how slice $S_k$ is scheduled at one cell impacts the channel quality (and hence scheduling outcomes) of another slice at a different cell. Our work deals with this challenge: how can we retain performance isolation while doing channel-aware scheduling across slices and interference management across cells.

As depicted in Fig. 2, network operators commonly scale cellular capacity by augmenting coverage regions of macro base stations with lower powered “small cells” [47, 9, 5]. These small cells can be deployed in locations where the macro cell provides poor coverage due to increased distance or obstacles such as high buildings. A macro-cell region (with coverage radius of 1-25Kms) can be expected to support 1-10 small cells [2] (each with coverage radius spanning 0.1-1Kms) [42, 7, 41, 13]. Recent years have witnessed a sharp increase in such multi-cell deployments. At the end of 2022, there were 142K macro-cell towers across the US, and a total of 452K outdoor small cell nodes [58]. It is projected that by 2027, there will be 13 million outdoor 5G small cell deployments at a global level [50].

One way to deploy multiple cells is to partition the frequency bandwidth among cells, such that they never interfere on the same band. However, such static partitioning of frequency can result in severe under-utilization of the network and reduced user throughput (as highlighted in Appendix A). Therefore, an increasingly common mode of deploying multiple cells is for all the cells to share the same frequency band, which avoids partitioning the bandwidth, thereby increasing the overall network capacity [20, 16]. This, however, results in interference, and necessitates a mechanism for managing it.

Interference management has been studied in 4G context (without slicing). CoMP (Coordinated Multi-Point) [1, 56, 39, 31, 38, 42] encompasses a suite of techniques for managing interference that vary in their deployment complexity, including techniques such as coordinated beam-forming, coordinated muting, etc. In this paper, we focus on coordinated muting, which enables muting selected RBGs on selected cells during each TTI to manage interference.

When compared to other CoMP strategies, muting is far more feasible to implement. It requires the cells to synchronize and coordinate with one another at per-TTI timescales of $\approx 100s\mu s$. Prior work demonstrated the feasibility of deploying muting [31, 32], and of achieving sub-ms coordination between cells [42]. More recently, the move towards virtualized RAN, where RAN processing is deployed in servers hosted in the edge datacenters (away from, but still in the geographical vicinity of, the base station) further provides a ready-made platform where such coordinated decisions across multiple cells can take place [37, 61], that can be leveraged by our system.

However, interference management (and muting) has only been explored in the context of optimizing a single objective (i.e. proportional fairness) across all users [4, 28, 48] – we provide a primer on how muting works in this traditional single objective scenario in §3.2. The focus of our work is to manage interference via muting in the presence of RAN slicing, with each slice optimizing its own diverse objective.

Summarizing the context so far, prior work on RAN slicing has not considered the problem of interference management in multi-cell scenarios. On the other hand, the prior work on interference management has not considered the new regime of network slicing where users are grouped into slices with diverse objectives.

One exception, that considers both interference management and slicing, is RSEP [17]. RSEP proposes an extreme solution of assigning a given slice exactly the same set of resource block groups (RBGs) at each cell. We depict this in Fig. 3 with the table on the left – slice $S_1$ is assigned the first RBG at each of the three cells ($C_1$, $C_2$ and $C_3$), while slices $S_2$ and $S_3$ are assigned the second and third RBG respectively at each cell. Such an assignment ensures that any muting decision for an RBG only affects users from the same slice at the neighboring cells. It thereby reduces the problem of muting to the no slicing (single objective) setting. However, such a constrained assignment inherently ignores channel diversity, where the channel quality on a given RBG varies drastically depending on which user is scheduled on it [12, 11, 59] (as explained in §2.1). This results in inefficient usage of the radio spectrum and reduced network throughput (up to 40% lower than channel-aware slicing [12]).

Fig.3 illustrates this through a simple example. The table on the left shows the schedule with RSEP, where each slice is restricted to use the same RBG at each cell. Slice $S_1$ prefers to mute its assigned RBG at cell $C_2$ to boost the channel quality of its user $U_{11}$ at the neighboring cell $C_1$. Slice $S_1$’s user at $C_3$ ($U_{13}$) is not significantly impacted by interference from $C_2$. The numbers in parenthesis denote the data rate (throughput) achieved by the scheduled user. The total data rate experienced by Slices $S_1$, $S_2$, and $S_3$ (summed across their scheduled users) is 60, 60, and 85 respectively.

The table on the right shows a more “flexible” assignment of RBGs across slices that can leverage channel diversity. Such an assignment can achieve superior performance by allowing different slices to use different RBGs across different cells, as per their respective channel qualities (e.g. RBG 2 is assigned to slice $S_2$ at cell $C_1$ and to slice $S_3$ at cell $C_2$). The unrestricted assignment further allows a given muting decision to simultaneously benefit multiple slices thus amortizing the cost of muting (for e.g. muting the transmission of RBG 1 at cell $C_2$ boosts the channel qualities of users $U_{11}$ and $U_{33}$ in Slice $S_1$ and Slice $S_3$ respectively). In this case, the total data rate, summed across all scheduled users for Slices $S_1$, $S_2$, and $S_3$ is 110, 110, and 140 respectively; which is higher across the board when compared to RSEP.

While this is a toy example, our evaluation with real-world traces in §6 reinforces the performance benefits of such “flexible” channel-aware assignment of RBGs across slices and cells, when compared to RSEP’s restricted assignment.

However, enabling such flexible channel-aware assignment requires reasoning about the muting decision across multiple slices with diverse objectives: Is it okay to mute the first RBG at $C_2$ to benefit slices $S_1$ and $S_3$? Which slice incurs the cost of this muted RBG and how? We need a way to answer such questions in a manner that still retains performance isolation across slices. Our system, RadioNinja, deals with this challenge.

We next detail how muting decisions are made without slicing where all users share a single global objective. Past work adopts a greedy strategy for this [4, 28, 48]. In each TTI, we greedily pick RBG $R_i$, and analyze the cost-benefit trade-off of muting that RBG at each cell, one cell at a time (referred to as the muting hypothesis). It is beneficial to mute $R_i$ at cell $C_j$ if the boost in channel quality at the neighboring cells overpowers the penalty of losing out on that RBG at $C_j$. To assess this, we compute the system-wide total score under each muting hypothesis $H_j$ (where $R_i$ is muted at cell $C_j$) and under no muting (denoted as hypothesis $H_0$). The total score under hypothesis $H_j$ is computed as: ${\sum }_{u_k\in U(R_i,H_j)}score(u_k,R_i)$. Here, $U(R_i,H_j)$ denotes the set of users that are scheduled on $R_i$ at each cell under hypothesis $H_j$. The user scores are computed as described in §3.1, as per the shared global objective.

The difference between the total score of a muting hypothesis $H_j$ and no muting hypothesis $H_0$ gives us the net benefit of muting RBG $R_i$ at cell $C_j$. If the net benefit is negative for all muting hypothesis, $R_i$ is not muted at any cell. Otherwise, we select the hypothesis with the largest net benefit and mute the RBG at that cell.

Fig. 4 presents an example for this, where we have 3 cells and want to determine whether or not to mute the transmission of a cell on a given RBG $R_i$. Hypothesis $H_0$ is when $R_i$ is not muted at any cell (top row in Fig. 4). The three cells schedule UEs $U_1$, $U_4$ and $U_7$, that do not experience significant interference from other cells. The total score for $H_0$ (last column) is computed by adding the scores of each of these three UEs (indicated in parentheses). The next hypothesis $H_1$ (second row) is where $R_i$ is muted at $C_1$ – this impacts the scheduling decision at cell $C_2$, boosting channel quality of “edge” UE $U_3$. The set of scheduled UEs, based on whom $H_1$’s total score is computed, therefore includes $U_3$ at $C_2$ and $U_7$ at $C_3$. Likewise, hypothesis $H_2$ (third row), where $R_i$ is muted at $C_2$, boosts the channel quality of “edge” UEs $U_2$ and $U_6$, and they get scheduled on this RB at cells $C_1$ and $C_3$ instead. The overall score of $H_2$ is computed by adding their scores. We similarly consider the impact of muting $R_i$ at $C_3$ and compute the overall score of $H_3$. Comparing scores of $H_1$, $H_2$ and $H_3$ with $H_0$, we can see how muting $R_i$ at $C_1$ or $C_2$ produces a positive net benefit while muting it at $C_3$ incurs a loss. The benefit of muting at $C_2$ is higher, so the system will mute $R_i$ at $C_2$. This freezes the muting and scheduling decision for that RBG, and we then move on to the next RBG and repeat the same steps.

A few points are worth noting:

• $\mathrm{■}$

  In the above example, we could have additional hypotheses where we mute $R_i$ on more than one cell (resulting in $2^M$ hypothesis for each RBG, where $M$ is the number of cells). We observed that muting multiple cells for the same RBG provides limited benefit since most interference occurs between the macro-cell and small cells, i.e., muting either the macro-cell or one of the small cells that interferes with the user being served by the macro-cell is sufficient. So we restrict the assessment to consider muting up to one cell per RBG (making the number of hypotheses linear in $M$).

• $\mathrm{■}$

  The muting hypothesis is evaluated without actually muting the cell – it uses knowledge about the channel quality each user will experience with and without interference from the neighboring cells. (The 3GPP standard specifies the use of specific “resource elements” to enable such interference measurement [3, 15], as detailed in Appendix B).

While it is straightforward to reason about muting decisions when all users share a single common objective (as described in §3.2), RadioNinja must deal with the challenge of handling diverse slice objectives in a manner that maintains performance isolation between slices. Fig. 5 illustrates the challenge by considering how an RBG $R_i$ is assigned to users across slices under no muting and when it is muted at cell $C_2$. When $R_i$ is not muted (hypothesis $H_0$, top row), users of slices $S_1$, $S_2$ and $S_3$ are scheduled on $R_i$ at cells $C_1$, $C_2$ and $C_3$, respectively. Muting $R_i$ at $C_2$ (hypothesis $H_2$, bottom row) changes the channel quality (and scores) of users in the neighboring cells $C_1$ and $C_3$, causing $R_i$ to be assigned to a different set of slices ($S_4$ at $C_1$ and $S_5$ at $C_3$). It therefore appears as if the decision to mute RBG $R_i$ at cell $C_2$ benefits slices $S_4$ and $S_5$, at the cost of slices $S_1$, $S_2$, and $S_3$. How do we make a systematic choice in such a scenario – is it okay to help some slices at the cost of others? Wouldn’t that violate performance isolation across slices? RadioNinja effectively avoids making such choices using the following insight:

We now explain how RadioNinja computes the benefit of a given muting hypothesis for a given slice in $𝒮$. Let us refer again to the example in Fig. 5, and consider how to compute the benefit of muting hypothesis $H_2$ (i.e muting of RBG $R_i$ at $C_2$) that slice $S_4$ enjoys at the neighboring cell $C_1$. At a high-level, this can be computed by comparing the score of $S_4$ at cell $C_1$ with and without muting $R_i$ at $C_2$ (with scores defined based on the slice’s objective, as detailed in §3.1). However, naively comparing the assignment of $R_i$ under $H_2$ with the no muting hypothesis $H_0$, it would seem that the performance of $S_4$ (which was not even assigned $R_i$ at any cell under $H_0$) is far better under $H_2$ (where it is assigned $R_i$ at $C_1$). This is a misleading assessment that overestimates the benefit of muting. This is because, by getting scheduled on $R_i$ at $C_1$ under hypothesis $H_2$, $S_4$ uses up a quota of 1 RBG at $C_1$. It will therefore be scheduled on 1 less RBG among the remaining $N-1$ RBGs in that TTI under $H_2$, when compared to the remaining $N-1$ RBGs under $H_0$.

In order to truly assess the benefit of muting, we need to compare how the RBG quota used by $S_4$ to schedule user $U_{41}$ on $R_i$ under $H_2$ compares to the use of that RBG quota under $H_0$. This ideally requires comparing the total score of $S_4$ across hypotheses $H_0$ and $H_2$, not just for RBG $R_i$, but for all of the remaining $N-1$ RBGs in that TTI, such that we can account for how that RBG quota is subsequently used under $H_0$. This implies that we need to run the scheduler to compute the hypothetical assignment of UEs on these remaining $N-1$ RBGs for each muting hypothesis that we wish to greedily assess for each RBG. As we show in §6, the O($N^2$) complexity induced by this is prohibitively expensive, given the tight TTI timescales at which scheduling and muting decisions must be made. So how do we work around this?

As mentioned before, we account for the cost of muting RBG $R_i$ at $C_j$ by subtracting $\frac{1}{|𝒮|}$RBGs at cell $C_j$ from the quota of each slice in set $𝒮$ (that benefit from the muting). Assessing this cost for a given slice $S_k$ requires comparing how $S_k$ is scheduled on remaining RBGs at $C_j$ with and without muting (i.e. with and without the quota reduction). This causes two hurdles. First, as discussed in §4.2, re-running the schedulers across all remaining RBGs for each muting hypothesis is prohibitively expensive. Second, how do we even assess the impact of fractional quota reductions? RBGs must be wholly allocated to a single user, and cannot be split across multiple users. When multiple slices have non-integral quotas, the scheduler allocates the last RBG in the TTI (that must ideally have been split between these slices) to one of these slices picked randomly. It then keeps track of the (fractional) surplus or deficit in quota allocation, that rolls over to the next TTI. Rolling over of fractional offsets ensures that a slice pays its due cost of muting over time. But the impact of that might not be seen in that TTI, and re-running the scheduler over multiple TTIs would increase the computational complexity further.

We use the benefit and cost computed as described above to assess whether the muting hypothesis is valid as outlined in §4.1. We empirically observed that in most cases, at most one muting hypothesis is valid for a given RBG (for the reason explained in §3.2). When there are multiple valid muting hypothesis $H_j$, we pick one as follows. For each valid muting hypothesis $H_j$, we compute the total benefit seen by each slice in set ${𝒮}_j$ (where set ${𝒮}_j$ is the converged set of benefiting slices in $H_j$) and divide it by the cost incurred by that slice. We then sum this benefit to cost ratio across all slices in set ${𝒮}_j$, and favor the muting hypothesis that has the highest value for this sum.

We simulate a multi-cell deployment comprising one macro cell and four small cells (similar to what is depicted in Fig. 2) [2, 42]. We configure each cell with a bandwidth of 100 MHz, split across 512 RBs per TTI, grouped into 64 RBGs [53]. The small cells are deployed uniformly within the macro-cell’s coverage region at a randomly-chosen distance of at least 300m from the macro-cell and 500m from neighboring small cells. We simulate 160 users split across 8 slices [2]. We situate each user within the coverage region of a randomly selected small cell with a probability of 2/3, and outside of the coverage region of any small cell (thereby attached to the macro-cell) with probability 1/3. Of the total users, 60% are static while 40% are mobile (traveling at speeds ranging from 3 to 30 km/h). Once a user is generated, we apply the 3GPP urban area path loss model to calculate the user’s wideband channel quality (aggregated across all RBGs) based on its distance from each cell. We then assign a real-world trace to each user that maps the computed wideband value for each cell (thereby extracting the effects of channel diversity from real-world conditions). We use the standard policy of assigning each user the serving cell from which it experiences highest wideband channel quality. We configure each slice to have the same quota of RBGs at each cell. We vary the slice objectives and user traffic patterns across different scenarios, and test the scenarios against different scheduling and muting schemes described above. For each setting, we run experiments with multiple random seeds that end up varying the specific user locations, and user distribution across cells and slices. We capture different slice objectives by varying the parameters of weighted $\alpha$-fairness (as described in §5). We experiment with two categories of user traffic patterns: (i) backlogged (representing high bandwidth video streaming, gaming, etc that would saturate the datarate allocated to the user), and (ii) web flows (fixed sized flows with Poisson inter-arrival times and flow sizes drawn from a heavy-tailed distribution [43], generated at an average rate of 10Mbps).

We use RadioSaber without muting as the reference baseline and report the relative changes induced by RadioNinja, SOM, and RSEP over it for the given slice-objective.

In Fig. 12(a) we show how RadioNinja (with its heuristics to assess benefit and cost of muting without repeatedly re-running the scheduler) compares against a variant that re-runs the schedulers for a more rigorous assessment (we refer to this variant as exhaustive assessment). We find that RadioNinja’s well-designed approximations produce no significant difference (+2.8%) in the outcome. For brevity, we only report a key result for scenario 3 – median FCT of prioritized users in WPF slices –(Fig. 12(a)), but trend holds more generally.

We use the software implementation (described in §5) on a single core to evaluate the runtime latency of both RadioNinja and the exhaustive assessment baseline. Fig. 12(b) shows the results with 5G numerology 1 configuration (500$\mu$s TTI and 32 RBGs), four small cells and five slices. We vary the total number of users from 200 to 800 (3GPP standards recommend 200-400 number of users for a cluster of 1 macro-cell and four small cells [2]). We find that RadioNinja’s runtime latency comfortably fits within the TTI budget with up to 600 users. The exhaustive assessment approach, in contrast, exceeds the TTI budget even with 200 users. The exhaustive approach can potentially be parallelized to use more cores and reduce latency, but the increased cost and power might not be worth it given that RadioNinja’s approximations work as well (as shown in Fig. 12(a)). We repeat these experiments for other 5G numerologies in Appendix D, confirming the same trends.