EcoCell: Energy Conservation Through Traffic Shaping in Cellular Radio Access Networks

Our first observation is that reducing the number of subframes, by packing more data in a single subframe, reduces the energy consumption of a base station. This is because transmitting every sub-frame induces additional control (DCI) transmissions. Given static channel conditions, we can pack more data in a subframe by utilizing more PRBs in a single subframe.

To validate this observation, we setup a microbenchmark. We use a server to transmit UDP data (200 Mb total) to a UE through the base station using different transmission patterns. We start by transmitting continuously at 2Mbps from the server. Then, we transmit at 10 Mbps for 0.2 seconds, followed by no transmissions for 0.8 seconds (i.e., 20% duty cycling). We repeat these experiments for 20 Mbps (0.1 seconds on period every 1 second) and 66 Mbps (0.03 seconds on period every 1 second). These transmission patterns achieve the same effective rate, with different levels of burstiness. We plot the PRB utilization per subframe for each of these flows in Fig. 6(b). As expected, with increasing burstiness, the PRB utilization goes up. Since our base station uses 20 MHz bandwidth, there are a total of 100 PRBs available for each subframe. We see a clear benefit of increased PRB utilization (i.e., more utilization per subframe, but fewer total subframes) in terms of energy savings in Fig. 6(a). Across 10 runs of these experiments, the average energy consumption drops by 53.2% for the same amount of data.

This experiment demonstrates that transmitting more PRBs per subframe (and hence fewer subframes) leads to energy savings because of much less control signaling overhead. This is validated by the corresponding reduction in the number of DCIs (in blue in Fig. 6(a)) by 57%. The energy consumption drops linearly with the drop in downlink control signaling.

We observe that the distribution of PRBs across UEs significantly influences energy usage. In a cellular network, the base station’s scheduling policy dictates the allocation of PRBs to multiple UEs in each subframe. For instance, consider two UEs requesting 100 PRBs each. One allocation strategy, (i), evenly divides the PRBs across subframes, providing each UE with 50 PRBs in two different subframes. Alternatively, (ii), one subframe could be entirely allocated to the first UE, and the next subframe entirely to the second UE. Our findings suggest that the latter approach leads to significantly reduced energy consumption.

To verify this, we saturate the cellular link with UDP traffic at 100Mbps (higher than base station capacity). In each experiment, we divide this rate equally into different numbers of UEs to compare the energy consumption. Saturating the link with multiple UEs causes the base station to share a subframe’s resource blocks with multiple UEs. We plot the energy consumption (in orange) in Fig. 8(a); it shows that when number of UEs increases from 1 to 5, the energy consumption increases by 14%. Fig. 8(b) shows the number of PRBs assigned to individual UE per subframe. When the number of UEs increases, fewer PRBs are assigned to each UE in each subframe, demonstrating sharing of the subframe by the multiple UEs. For instance, with a single UE, the median PRB per UE is 100, whereas with four UEs, it drops to 28 due to sharing. These two results combined show the implication that: more UEs shared per subframe leads to more energy consumption.

Including multiple UEs in a single subframe requires control information for each UE in that subframe. Therefore, more UEs sharing a subframe leads to more control signaling overhead, as shown in Fig. 8(a) blue. The number of DCIs increases by 59% when the number of UEs increases to 5. Considering the allocation strategy examples mentioned above, there will be 4 DCIs in total (2 in each subframe for 2 UEs) in the strategy (i), while only 2 in strategy (ii), each in a subframe serving a single UE.

The modulation and coding scheme, or MCS, governs how much user data can be encoded into each resource block. The base station computes per-UE MCS values based on the UE’s downlink signal quality [2]. The base station energy consumption increases with decreasing MCS.

We vary the MCS value by placing one UE at nine different locations shown on the yellow dots in Fig. 7(a), each separated by around 40 meters; The red star shows the BS’s location. At each location, we ensure that the base station is not serving any other users, and then run an iperf TCP flow with a fixed throughput of 25 Mbps to download data through the base station for 40 seconds. We plot the aggregated MCS values by adding MCS values from both spatial streams and the result is shown in Fig. 7(b). Different background colors from left to right represent UE moving away from the base station. As the UE moves away from the BS, its MCS decreases, and the BS’s power draw increases.

As the MCS decreases, the base station needs to use more frequency and time resources (i.e., subframes and PRBs) to exchange the same amount of user data, increasing the BS energy consumption. We plot the number of non-empty subframes to verify this relationship in Fig. 7(c).

To validate that subframes are under-utilized in commercial BSes, we sniff traffic from a campus base station and a base station in a large city, using the setup described in Sec. 4). Figure 9(d) shows that over half of the subframes use fewer than 20% of the PRBs on campus and fewer than 29% of the PRBs in a major U.S. city downtown, showing high subframe underutilization and the potential to reduce energy consumption. Our measurements match the well-known observation that cellular networks are generally under-utilized [22].

We conduct the following microbenchmark to show how EcoCell can improve subframe packing. We attach one UE to our BS, and start a 5 Mbps UDP-based iperf download. We plot the on-air throughput observed using NG-Scope in Fig. 9(a). With EcoCell configured with $t_{\text{burst}}=0.1s$, the traffic pattern has large bursts every 100 ms, as expected.

Figure 9(b) shows that increasing $t_{\text{burst}}$ improves subframe packing, at the cost of higher added latency (assuming packets arrive uniformly, the added latency is $t_{\text{burst}}/2$). For example, for a 5 Mbps iperf download without EcoCell, non-empty subframes use fewer than 20% of their RBs on average. In contrast, bursting at 50 ms intervals increases the average RB utilization of non-empty subframes to over 60%. Figure 9(c) shows the corresponding reduction in BS’s dynamic energy used for the download: 137J without EcoCell, and 93J with 50 ms burstiness. We note that we did not observe significant subframe packing benefits in iperf TCP flows which were naturally bursty. However, our experiments (Sec. 7) show that subframe packing saves power in other important TCP workloads such as video streaming over HTTP.

How should a cellular operator choose an appropriate burst interval $t_{\text{burst}}$? We believe that it depends on the energy-latency tradeoff that the operator is willing to make, which may vary across different cell sites and time of day. Operators already make such energy-performance tradeoffs with energy efficiency optimizations like turning off cell sites or carriers at certain hours of the day (e.g., post-midnight). We explore this tradeoff in detail in Sec. 7 for various applications. Based on our evaluation results, operators can design custom, application-adaptive algorithms in RAN Intelligent Controller(RIC) to directly provide millisecond-level bursty pattern adjustment. For example, URLLC traffic (such as drone, remote surgery), which requires ultra-low latency (<10 ms), can be assigned a small bursty period or be allowed to pass directly, while eMBB services could use a larger bursty period simultaneously. Empirically, we observe that $t_{\text{burst}}=50$ ms works well for all studied applications with significant energy gains and negligible impact on application performance.

How often are subframes shared across UEs and how many UEs share a subframe in practice? We analyze this question in our week-long trace from two commercial base stations as described before (Sec. 4). We plot the percentage of subframes shared by more than one UE in Fig. 11(a). Over 20% of the subframes are shared by two or more UEs during the day because of high traffic demand. While there is less sharing at night, the ratio remains above 5% on campus and 10% in the city. These patterns show a significant opportunity for using UE segregation to reduce subframe sharing and save energy.

To show the effect of UE segregation, we conduct a microbenchmark experiment shown in Fig. 11(b). We start by saturating the cellular link with 5 UEs each receiving UDP traffic at 20 Mbps. We find that over 60% of the subframes are shared by multiple UEs (blue line). We enable our UE segregation module 120 seconds into the experiment, which reduces the fraction of shared subframes to near-zero. Correspondingly, the BS’s dynamic power draw decreases from roughly 8 W to 4 W.

The challenge with an RSRP-based temporal shifting algorithm lies in the possibility that future RSRP values may deteriorate, resulting in unnecessary traffic delays and no energy savings. To address this challenge, we adopt a heuristic-based approach. We make decisions regarding traffic delays and their durations based on observed average RSRP values. Specifically, we adjust the throughput of each flow every three seconds in proportion to the observed RSRP. This adjustment is linear, ensuring that higher RSRP values correspond to increased throughput and vice versa.¹¹1In practice, the MCS (and hence the datarate) is mapped by CQI (Channel Quality Indicator), we use RSRP instead due to the data availability and its high correlation with CQI. We place two additional constraints on the throughput. First, throughput adjustments should not violate any deadlines, such as flow completion times in long-running flows (when they are available). Second, we scale the overall throughput deviation to be between the minimum and maximum throughput supported by the base station. In practice, without access to the RAN, we implement such traffic shaping through EcoCell’s TUN interface.

We present the MCS distribution of devices observed in our commercial traffic trace in Fig. 12(c). The distribution showcases a broad range of MCS values, with a slight concentration observed between 10 to 16. This MCS landscape underscores the significant potential for generalizing our energy savings solution via temporal shifting in commercial cellular networks.

To demonstrate the impact of temporal shifting, we conduct a real-world experiment and present the results in Fig. 12. In this experiment, a user device attempts to download a 200MB file from the server, with the server download speed set at 20Mbps, We set the desired job finishing time to be 600s.

Fig. 12(a) illustrates the download process without EcoCell, where the client throughput remains constant despite fluctuations in RSRP values. For instance, between the 40 to 60-second interval, the RSRP increases from -110dBm to -90dBm, prompting a corresponding rise in transmission MCS (aggregated by 2 spatial streams) from 20 to 31. However, due to the constant throughput, energy consumption decreases by 62%. Conversely, during the 20 to 40-second interval, energy consumption is higher due to downloading at a constant throughput during a period of low RSRP.

In contrast, Fig. 12(b) demonstrates our dynamic throughput control mechanism. As RSRP changes, EcoCell adjusts the throughput accordingly. For instance, during the 10 to 50-second interval, with RSRP at -100dBm, EcoCell throttles the traffic to a lower throughput of 2.6Mbps, awaiting improved RSRP conditions. Overall, our method reduces energy consumption during file downloading from 362J to 318J, a reduction of 12%, while ensuring the file is downloaded within the specified time frame.

How do we identify delay-tolerant traffic? Such traffic classification has been explored in past work extensively [57, 65]. In general, such traffic classification is also increasingly available to operators through slicing indicators [35, 6]. Therefore, for our work, we assume that delay-tolerant traffic characteristics are available to us either through classification algorithms or through dedicated delay-tolerant slices.

We use Google Meet as a representative video conferencing application [16]. Video conferencing is an extremely delay-sensitive application. We set up a meeting room using Chrome on the server with a virtual camera which is fed with a pre-recorded video. A client joins the meeting through the cellular network with the same virtual camera setting as well. We use the WebRTC extension in Chrome to obtain the QoE metrics including frame per second, video resolution, etc.

Live streaming applications such as live games, TV, news, and events have more delay sensitivity than YouTube/Netflix but less urgency than Google Meet. To emulate such workloads, we stream the "EnvivioDash3" video using DASH from a server to the UE. This video is 193 seconds long with 48 chunks and has been commonly used as a benchmark for past work [43, 49]. The server performs Adaptive Bitrate (ABR) adaptation using the robustMPC algorithm [64] to dynamically select the video bitrate among {300, 750, 1200, 1850, 2850, 4300} kbps. The client player is a Google Chromium browser and we set the DASH playback buffer capacity to be 2 seconds in order to enforce the real-time effect. We achieve this by setting the dash.js parameters in the html: setlivedelay() and setStableBufferTime().

Video streaming applications such as Netflix and YouTube account for a large fraction of network traffic. We adopt the same setup as above but we set the DASH playback buffer capacity to be 60 seconds [49] since it has less urgency.

The client downloads websites from the HTTP server through the cellular network. We choose 10 popular websites and obtain not only the HTML but CSS and multimedia assets as well and put them onto the server. The average website size is 6.5MB with the highest of 17.9MB and the lowest of 2.4MB. The client mimics the user browsing these websites with intervals of 5 seconds.

We transfer a 400MB dummy file from the server. The client uses the wget tool to download from it through the cellular network.