5G MEC Computation Handoff for Mobile Augmented Reality
Pengyuan Zhou, Benjamin Finley, Xuebing Li, Sasu Tarkoma, Jussi Kangasharju, Mostafa Ammar, Pan Hui
55G MEC Computation Handoff for Mobile Augmented Reality
Pengyuan Zhou
University of [email protected]
Benjamin Finley
University of [email protected]
Xuebing Li
Aalto [email protected]
Sasu Tarkoma
University of [email protected]
Jussi Kangasharju
University of [email protected]
Mostafa Ammar
Georgia Institute of [email protected]
Pan Hui
University of HelsinkiHong Kong University of Science [email protected]
ABSTRACT
The combination of 5G and Multi-access Edge Computing (MEC)can significantly reduce application delay by lowering transmis-sion delay and bringing computational capabilities closer to theend user. Therefore, 5G MEC could enable excellent user experi-ence in applications like Mobile Augmented Reality (MAR), whichare computation-intensive, and delay and jitter-sensitive. However,existing 5G handoff algorithms often do not consider the computa-tional load of MEC servers, are too complex for real-time execution,or do not integrate easily with the standard protocol stack. Thusthey can impair the performance of 5G MEC.To address this gap, we propose
Comp-HO , a handoff algorithmthat finds a local solution to the joint problem of optimizing signalstrength and computational load. Additionally,
Comp-HO can easilybe integrated into current LTE and 5G base stations thanks toits simplicity and standard-friendly deployability. Specifically, weevaluate
Comp-HO through a custom
NS-3 simulator which wecalibrate via MAR prototype measurements from a real-world 5Gtestbed. We simulate both
Comp-HO and several classic handoffalgorithms. The results show that, even without a global optimum,the proposed algorithm still significantly reduces the number oflarge delays, caused by congestion at MECs, at the expense of asmall increase in transmission delay.
Cellular networks are a vital part of modern society with novelnetwork technologies enabling an expanding array of use casesfrom simple NB-IoT sensors to immersive mobile virtual reality.Fifth-generation mobile networks (5G) specifically provide supportfor much higher frequencies (up to 52.6 Ghz) with larger bandwidths(up to 400 Mhz) and lower radio access network delay (around 10 ms)in comparison to LTE.Relatedly, Multi-access Edge Computing (MEC), another novelnetworking technology, supports deploying compute nodes nearexisting network nodes in the mobile network structure (often asservers co-located with base stations in the radio access network).Thus user applications that require computation can lower totaldelay by sending the computation request to a physically and hop-wise closer compute node rather than to a remote cloud server [34, 45]. MEC is particularly suitable for computation-intensive anddelay and jitter -sensitive applications such as mobile augmentedreality (MAR) [11, 21, 43].MEC has recently been standardized by ETSI thus detailing thepotential for the tight integration of MEC and 5G technologies. Insuch a context, the handoff process [37] between base stations isan important and growing concern. This is because such a handoffoften means the user will also be served by a different MEC server.Under the assumption that different MEC servers have varying ca-pabilities and loads, the user application performance will thus beaffected by the capability and load of the new MEC server. However,current handoff decisions are typically based primarily on commu-nication measurements such as signal strength, without concern forthe status of MEC servers. Therefore, a handoff that improves signalstrength could still reduce overall application quality due to loaddifferences of MEC servers. This concern is especially important in5G networks given their typically small cell sizes (<500m) whichimplies more frequent handoffs. We denote this issue as the
MECHO problem.To address this issue, in this work we propose
Comp-HO, a low-complexity, stand-friendly handoff algorithm jointly considering re-ceived signal quality from base stations and the computational loads ofthe co-located MEC servers.
In other words, when the signal strengthdegrades sufficiently or the serving MEC server is sufficiently over-loaded, the base station initiates a handoff and re-assigns the com-munication and computation processes to another base station andMEC server. We focus specifically on the MAR use case, which hasstrict real-time quality requirements. Furthermore, we show thatour approach strikes a good balance between the signal strengthand computational load concerns with large numbers of MAR users.Specifically, we first develop a MAR prototype and deploy theprototype in a 5G MEC testbed (shown in Figure 1) to measurebaseline performance with a traditional handoff algorithm. Wethen utilize the measurement results to drive a custom MEC-enabled
NS-3 simulator and compare the
Comp-HO algorithm to classichandoff algorithms on larger scale network simulations. Unfortunately, the commercial 5G base station does not allow reprogramming so wecannot test the
Comp-HO algorithm in the testbed. a r X i v : . [ c s . N I] J a n Zhou, et al.
Base StationOptical Fiber(Median RTT: 0.5 ms)Core Network MEC
Figure 1: The 5G testbed is composed of the core networkand the base station, which are interconnected via an opti-cal fiber. The core network hosts the MEC via its NG6 inter-face, which is designed for connecting to the data network.The base station has two 5G antennas and one LTE antenna(middle).
To the best of our knowledge, ours is one of the first efforts toaddress the 5G
MEC HO problem. Our contributions are threefold:(1) Defining the
MEC HO problem in 5G and proposing an al-gorithmic solution,
Comp-HO , that jointly considers signalstrength and computational load.
Comp-HO is simple andstandard-friendly and outperforms traditional algorithmswith only a small transmission overhead.(2) Measuring baseline performance with a MAR prototype in areal-world 5G MEC testbed with a traditional handoff algo-rithm. This baseline helps guide the parameter selection forthe custom
NS-3 simulations.(3) Carrying out reproducible and measurement-based simula-tions to evaluate
Comp-HO at scale using a custom MEC-enabled
NS-3 network simulator [4].The remainder of this paper is organized as follows. Section 2presents related work in the areas of edge offloading and handoffalgorithms. We describe mathematically the
MEC HO problem inSection 3 and propose
Comp-HO algorithm in Section 4. Section 5details the baseline measurements with the 5G testbed. Section 6presents the simulation setup and results. We discuss the potentialfuture directions of the work in Section 7 and conclude in Section 8.
Handoffs can be classified into horizontal and vertical [32, 35], orhard and soft [10, 36]. In this work, we focus on horizontal hardhandoff since we target a pure 5G networking environment andhard handoff is more common with LTE [16].Traditional works largely make handoff decisions based on UEmeasurements consisting of signal quality indicators such as re-ceived signal strength [26, 39], signal-to-interference-plus-noise ra-tio (SINR) [12, 40] and reference signal received quality (RSRQ) [19,41]. However, given the novelty of MEC, the mentioned algorithmsonly focus on transmission metrics without taking edge computa-tion into consideration. The lack of such algorithms could be partlybecause the radio access network delay of LTE does not supportsome MEC-assisted applications [15] compared to novel standardslike 5G, therefore lessening the motivation for LTE + MEC research. As MEC and 5G techniques evolve, there has been some MEC-awarehandoff research. Nasrin et al. [27] propose a handoff algorithm thatjointly considers signal quality and computational loads. Sardellittiet al. [33] and Mao et al. [23] focus on the joint optimization ofradio and computational resources considering energy consumptionand user experienced delay. Basic et al. [5] propose a fuzzy logichandoff algorithm that selects a target node based on bandwidth,processor, and delay parameters of edge servers. Emara et al. [13]and Li et al. [20] both propose to improve handoff algorithms byconsidering MEC load in 5G heterogeneous networks and cloudradio access networks, respectively. Finally, Zhang et al. [46] detaila UE-based offloading algorithm for dense networks with MECsthat considers transmission delay, processing delay, and an energyconstraint under information uncertainty. Though in contrast toour work, they consider the task of deciding where to offload tasksgiven a single UE moving through multiple cells with MECs ofvarying capacities.Ma et al. [22] propose to build an efficient service handoff systemacross edge servers based on Docker container migration. Wang etal. [38] utilize a lightweight heuristic algorithm to reduce offloadingtask execution delay by jointly considering task information, smallbase station and user mobility information. Yu et al. [42] propose adynamic algorithm for partial offloading based on short-term mo-bility prediction to minimize energy consumption while satisfyingdelay requirements.To summarize, we find the existing MEC-aware handoff works fall short in several respects:(1) Most solutions do not provide realistic algorithms that takethe X2 application protocol into consideration [5, 13, 20, 22,27, 38, 42].(2) Some of the proposed algorithms have computational com-plexities larger than ๐ ( ๐ ) [38, 42]. In cases with larger num-bers of UEs, the complexity becomes problematic for realtime operation.(3) Related proposals tend to require additional message trans-missions between UE and base station to collect informationfor the handoff algorithm. This overhead impacts systemperformance through additional link transmissions and in-formation collection delay.(4) Related works do not inform their simulations with empirical5G measurements thus making their interpretations lessreliable [5, 13, 20, 22, 27, 38, 42].(5) Most related works lack detailed simulations. The relatedworks conduct only numerical modeling or simplified simula-tions without millisecond granularity and packet-level/multi-layer detail (e.g., from physical to application layer), thusexcluding some detailed dynamics only visible with suchsimulations [5, 13, 20, 27, 38, 42]. We address these shortfalls by proposing an easily-deployable hand-off algorithm with minimum overhead. We code the key metric valuescollected from a real 5G testbed together with Comp-HO algorithminto an open source custom MEC-enabled
NS-3 simulator followingbase station protocol standard and perform a packet-level simulation. In this work, we define user experienced delay as the time between when user sendsa request and receives a reply.
G MEC Computation Handoff for Mobile Augmented Reality ,
Table 1: Notation Table ๐ข UE ๐ข๐ MEC ๐ S Serving base station and MEC server T Targeting base station and MEC server ๐ ๐ Processing queue length in MEC ๐๐ท ๐๐ข Transmission delay from ๐ข to ๐๐ ๐๐ข Signal quality for ๐ข received from the cell with ๐๐ ๐ข UE measurement: โจ S m ๐ข , ๐ด โฉ ๐ RSRQ threshold ๐ฟ Handoff offset
This section formulates the
MEC HO problem in the 5G contextwith MEC servers co-located with 5G base stations. The formula-tion focuses on optimizing the performance of MEC-assisted UEapplications with respect to experienced delay. Also we assumethe transmission delay between a MEC server and its co-locatedbase station is negligible (as also shown in the 5G testbed mea-surements, see Figure 1). We also note that there are many MEClocation deployment schemes, e.g., co-located BS and MEC or sev-eral BSs sharing MEC co-located with an MME. Since comp-HOcan be easily generalized to work these with different deployments(by considering MEC load only for HOs between BSs with differentMECs), we only focus on the co-located BS with MEC case.We let M โ { , , ..., ๐ } denote the ๐ MEC servers and U = { , , ..., ๐ } the U mobile UEs in the system. Each MEC server has afixed capacity ๐ (maximum queue length) and is connected directlyto a 5G base station via fixed-line Ethernet. The time horizon isdiscretized into slots of equal periods indexed by ๐ก โ N . We noteseveral important MEC server assumptions:(1) Homogeneity of MEC servers : The MEC servers have the samedata processing rates.(2)
Job-level migration : The tasks of a job may execute on anygiven MEC server.(3)
Task atomicity : A task cannot be split across MEC servers.(4)
Non-preemptive task scheduling : A task being processed can-not be interrupted by any other task.Let ๐ท ๐๐ข denote the transmission delay from UE ๐ข to MEC ๐ excluding the processing delay. In other words, ๐ท ๐๐ข consists of theuplink and downlink delay. Normally, either the uplink or downlinkhas larger data packet sizes and thus dominates the transmissiondelay. For example, uplink delay dominates the transmission delay,since the uplink packets to MEC servers contain much more datathan the downlink packets for most MAR offloading applications.To simplify the problem, we consider only the dominant direction ofdata transmission, ๐ท ๐๐ข , which solely depends on the signal qualityreceived by ๐ข from ๐ , i.e., ๐ ๐๐ข : ๐ท ๐๐ข = ๐ ( ๐ ๐๐ข ) . (1)where the function ๐ : ๐ โ ๐ is monotonically decreasing.Let ๐ ๐ denote the processing queue length of MEC ๐ at a pointin time. We assume the change in user experienced delay duringthe handoff from MEC S to T depends on the difference in signal Algorithm 1:
Comp-HO algorithm parameter : ๐ ๐๐ข โ RSRQ ๐ข received from ๐๐ ๐ โ Max queuing time in ๐ UE Measurementthread
ReportUeMeasurment( ๐ ๐ข ) : while ๐ข offloading to ๐ do S m ๐ข โ ๐ ๐๐ข for all probeable MECs, m โ M ๐ ๐ข โ โจ S m ๐ข , ๐ด โฉ // ๐ด โ > App info sendUeMeasurement ( ๐ ๐ข ) Handoffthread updateUeMeasurement( R u ) : R u โ ๐ ๐ข from all connected UEs, u โ U thread updateLoad( Q m ) : Q m , A โ โจ ๐ ๐ , ๐ด โฉ from all nearby MECs, m โ M thread Hand off u : if ๐ S ๐ข < ๐ then // ๐ -> RSRQ threshold ๐น ( ๐ T ๐ข , ๐ T ) = max_element( ๐น ( S m ๐ข , Q m )) if ๐น ( ๐ T ๐ข , ๐ T ) โ ๐น ( ๐ S ๐ข , ๐ S ) > ๐ฟ then // ๐ฟ -> Handoff offset SendHoRequest ( T )qualities and queue lengths as follows: โณ ๐ท S , T ๐ข = ๐ (โณ ๐ S , T ๐ข , โณ ๐ S , T ) , (2)where โณ ๐ S , T ๐ข = ๐ T ๐ข โ ๐ S ๐ข . The function ๐ denotes that both signalstrength and computational load are considered.Let ๐ S , T ๐ข indicate whether MEC S hands off ๐ข to T as follows: ๐ S , T ๐ข = (cid:40) S hands off task of ๐ข to T , . (3)We can then formulate an optimization problem as follows,min โ๏ธ ๐ข โU โณ ๐ท S , T ๐ข ๐ S , T ๐ข s.t. S , T โ M , S โ T (4) Local vs Global : The problem aims at optimizing user experienceby minimizing overall user experienced delay for all UEs. The prob-lem can be seen as a linear sum assignment problem which is alsoknown as a minimum weight matching in bipartite graphs. Bal-anced assignment algorithms such as the Hungarian algorithm [9]can be used to solve this problem. Therefore, an optimal solutionto the problem with global information is possible.However, a global solution faces three challenges: 1) the potentialfor a single point of failure (the node making the global decisions), 2)the delay overhead caused by global information collection and de-cision dissemination can degrade system performance, and 3) mostreal-world standards require base stations to make their own hand-off decisions (thus modification of standards would be required).Therefore, we instead develop
Comp-HO as a local optimizationalgorithm and leave global optimization for future work.
Zhou, et al.
UE TargetgNB SourcegNB AMF(s) UPF
Xn HO RequestXn HO Request ACKUE Measurement ReportHO CommandXn SN Transfer StatusHO Con๏ฌrm N2 Path Switch RequestN2ย Path Switch ACK Modify PDUSessionsXn UE Context Release
TargetMEC SourceMEC
5G NEF or RNI Noti๏ฌcation (MEC Load Info)5G NEF or RNI Noti๏ฌcation (MEC Load Info)Xn (MEC Load Info) 5G NEF or RNI Cell Change Noti๏ฌcation User Context
Comp-HO
Figure 2: High-level
Comp-HO flow diagram using 3GPP 5G and ETSI MEC [14] terminology.
Comp-HO specific flows aremarked in blue.
In this section, we first describe the Comp-HO handoff algorithm.Then we illustrate the integration of Comp-HO into the existingESTI MEC standard information flows. Finally, we discuss the scal-ability of Comp-HO.
To perform a
MEC HO , each UE sends measurement reports to theserving base station. The serving base station runs the handoffalgorithm and decides when to initialize a handoff to a target basestation. We first describe the UE measurement and then the hand-off algorithm. Table 1 summarizes the related notations. Each UEcollects the signal qualities (RSRQ) from all nearby base stationsand sends them to the serving base station together with the infor-mation of its MEC-assisted applications in the report as shown online 4 in Algorithm 1.In parallel, each base station collects UE measurements fromconnected UEs and load information from nearby MEC servers,respectively (line 5 to 6). The load information includes metadataof MEC-assisted applications running in each MEC server and theprocessing queue lengths. ๐น ( ๐ ๐๐ข , ๐ ๐ ) denotes the weighted sumof signal quality and queue length to perform the optimization. Tokeep the complexity low, ๐น () follows a linear form: ๐น ( ๐ ๐๐ข , ๐ ๐ ) = ๐ค ๐ โ ๐ ๐๐ข โ ๐ค ๐ โ ๐ ๐ . In the simulation, we iteratively tried differ-ent sets of weights and offsets to optimize performance. The basestation starts the handoff of a UE if its RSRQ fails to meet thethreshold (line 7). Utilizing the collected load information and sig-nal quality values, the base station selects the target base station and MEC server and sends the handoff request (line 8-10). The offsetmetric, ๐ฟ , is introduced to avoid the ping-pong effect (line 9). In terms of interaction between 5G network elements, Figure 2illustrates the high-level message flows between such elements justbefore and during a
Comp-HO . The figure uses the terminologyand interfaces following standard 3GPP 5G and ESTI MEC [14],except that the Network Exposure Function (NEF) or Radio Net-work Interface (RNI) should allow passing load information backto the gNB/ng-eNB to be used in the
Comp-HO algorithm. TheseNEF/RNI interfaces are currently designed to provide radio networkinformation such as cell change notifications to the MEC systemfor user context (e.g., virtual machine) migration between MECservers.Also, we note that the analogous flows with a 5G radio accessnetwork and LTE EPC (like in testbed network) would be onlyslightly different in functions and terminologies. Overall, the flowsillustrate how
Comp-HO could potentially function after integratedinto current mobile standards.
The local algorithm is ๐ ( ๐ ) in time complexity where ๐ is the num-ber of sector UEs. This complexity is the same as the baselines (sin-gle signal indicator based handoff algorithms), so the algorithmscales in time at least as well as those. Whereas in terms of measure-ment messages, in our scenario the UE does not deal with MEC loadinfo messages directly but these messages are transferred betweenthe MECs and base stations over Xn links (as Figure 2 illustrates). G MEC Computation Handoff for Mobile Augmented Reality ,
Thus the messaging complexity scales with the number of MECsand base stations rather than UEs, thus allowing good scaling asthe number of UEs increase.
Our 5G MEC testbed is a 5G micro-operator [24] network built by ajoint national effort of academic and industrial partners. As shownin Figure 1, the testbed is composed of two parts: a 5G Core [1]and a base station. The core network is deployed in non-standalone(NSA) mode and the network functions (NFs) are implemented inLinux virtual machines located on servers in a single universityserver room. The MEC server is located in the same server roomand connected via Ethernet to the core network switches. The basestation is located on the roof of another building. The base stationhas two antennas for 5G and one for LTE, providing coverage overthe campus area. The base station is connected to the core networkvia optical fiber. According to our measurements, the median RTTbetween the base station and MEC is 0.5 ms.We develop a MAR prototype by running a custom Android clientapp on a Huawei Mate 30 Pro 5G smartphone (the UE) and a Linuxserver app on the MEC equipped with 8-core Xeon CPU, 16 GBmemory and Quadro K2200 GPU. The client app captures cameraframes at 10 frame rate (FPS). Then, it downscales the frames to480 ร
320 pixels and sends to the MEC.The MEC receives the frames and uses YOLO [31] to performobject detection. The object detection result is composed of a set ofbounding boxes of the detected objects as well as the object classesand detection confidences. Once the objects are detected, the resultis sent back to the UE and rendered on the screen, annotating theobjects from camera view.Unfortunately, due to technical and licensing limitations, wecannot implement and deploy the
Comp-HO algorithm into the basestation. Instead, we conduct network measurements to observe thebaseline performance of the MAR prototype system in Section 5.2.Then, in Section 6, we use these measurement results to inform thesimulation parameters thus helping to mitigate the gap between areal-world 5G network and the
NS-3 simulator.
We record the transmission delays of the frames and detection re-sults by timestamping during the sending and receiving on the UEand MEC. We connect the UE and the MEC via Ethernet before-hand to estimate the clock drift (within a confidence interval of<1 ms). We also select an off-peak time to conduct the measure-ments. The results, therefore, illustrate MAR performance with 5GMEC without non-MAR loads from other UEs.We walk along a fixed route during the measurement and collectthe results shown in Figure 3 and Table 2. Overall, the UE sends 540frames to the MEC over a period of 54 seconds. The median frame(uplink) transmission delay and the result (downlink) transmissiondelay are 32.0 ms and 2.0 ms, respectively. This large difference isdue to two primary reasons.(1) Firstly, the uplink bandwidth is much smaller than the down-link. According to our measurements, the downlink through-put is about 360 Mbps while the uplink only 30 Mbps. F r a m e T r an s m i ss i on D e l a y ( m s ) D e t e c t i on R e s u l t T r an s m i ss i on D e l a y ( m s ) Figure 3: Frame transmission delay and detection resulttransmission delay between the UE and the MEC, measuredwith the MAR prototype in 5G MEC testbed.Table 2: Measurement statistics of Figure 3. UE โ MEC:frame transmission (uplink). MEC โ UE: detection resulttransmission (downlink). UE โ MEC MEC โ UEMedian delay (ms) 32.00 2.00Jitter (ms) 9.5 0Packet loss (%) 0 .
06 0 . We next perform simulations to estimate the performance of the
Comp-HO algorithm at scale. The simulation setup aligns with someimportant metrics and results taken from the 5G measurements suchas the frequency, UE speed and lower bound of user experienceddelay.
Simulator:
We modify the existing LTE module of NS-3 so thateach eNB is co-located with three MEC servers (one for each sector)and any UE data packets are forwarded to a MEC server rather thanto a packet gateway (Figure 4). We also integrate the
Comp-HO algorithm into the LTE module. Each MEC server contains severalqueues that each process packets at a fixed rate. For a given MECserver, the processing queue length reported to the eNB (and thus tothe handoff algorithm) is the minimum length out of these queues. We use the LTE module rather than a 5G
NS-3 module because handoff support insuch modules [25, 29] is still limited.
Zhou, et al.
NS-3 eNBeNBeNB MECMECMECAPP Comp-HOSAP
Dependency Connection
Figure 4: Custom
NS-3 simulatorTransmission Setup:
For simplicity, each UE sends UDP packetsat a fixed FPS (20 Hz) but with random starting times to avoidsynchronization. As mentioned, the packets are forwarded by theeNB to the corresponding MEC for processing and after processingsent back to the UE. If after processing, the UE is no longer beingserved by the eNB that received the packet originally (for examplebecause of handoff) the packet is discarded. We denote this typeof packet loss as MEC-mobility packet loss . For performancemonitoring, the round-trip packet delay (which we refer to as theUE experienced delay) for each packet is recorded at the UE. Wealso track the number of handoffs. Variations:
We perform both basic simulations (with the notedparameters) and several simulation variations as follows.(1)
Handoff rate:
We vary the UE speeds to alter the effectivehandoff rate.(2)
FPS:
We vary the FPS to uniformly alter the MEC loads (giventhe fixed processing rate),(3)
Mobility:
We use two different UE mobility models (one witha center bias and one without) to illustrate the impact ofdifferent spatial UE distributions (e.g., a central crowd ofUEs).For further clarification on the mobility models, we use ran-dom waypoint as the baseline model as random waypoint has acenter bias [6] thus allowing a higher central user density and het-erogeneous load across MECs. Such heterogeneity naturally helpsillustrate the benefits of the
Comp-HO algorithm. However, for ref-erence, we also use a Gauss-Markov mobility model which does nothave such a center bias [8], and thus has a more homogeneous userdistribution and load across MECs. We also perform each simulationthree times (with different random seeds) to ensure performancedifferences are not due to random variation.
Benchmark:
As pointed out in section 2, most MEC-aware hand-off proposals do not take the X2 protocol into consideration [5, 13,20, 22, 27, 38, 42] and thus are incompatible with multilayer sim-ulators such as
NS-3 , or have significant complexity (larger than This loss is a challenge for MEC systems with LTE as the system cannot easily impactLTE traffic routing, thus forcing somewhat complex solutions. Luckily, 5G has flexibleuser plane functions that the external MEC system can change, thus reducing the issue.In any case, even removing this type of loss completely does not qualitatively changeour delay results. -200 Y ( m e t e r s ) X (meters) -10-5 S I NR ( d B ) Figure 5:
NS-3 simulation network layout map with SINRTable 3:
NS-3 simulation parameters
Parameter ValueNumber of base stations 18 Tri-sector eNBsLayout HexagonalIntersite Distance 350 mCenter Frequency 3.55 Ghz (band 22)Bandwidth 20 MhzPath Loss Model ITU-R P.1411 LoS [17]Height eNB: 45 m, UE: 1.5 mNumber UEs 50UE Mobility Random WaypointUE Velocity 2 m/s (7.2 km/h)Queues per MEC Server 1Simulation Area 1800x1300 mSimulation Time 30 s
Table 4: Benchmark Parameters
Algorithm Parameters
A2-A4-RSRQ
ServingRsrqThreshold=30 [2]NeighbourRsrqOffset=1 [2]
A3-RSRP
TimeToTrigger=256 (ms) [2]Hysteresis=3 (dB) [2] ๐ ( ๐ ) [38, 42]) and thus may require further improvements forpractical use. Additionally, to the best of our knowledge there areno available open source implementations of MEC-aware handoffalgorithms (besides ours [4]). Therefore, we leave comparisons withother MEC-aware handoff algorithms for future work when moreimplementations are available.Instead, we compare the Comp-HO algorithm to two existingLTE handoff algorithms (
A2-A4-RSRQ and
A3-RSRP ) and to a sce-nario with no handoffs (NoHO). The
A2-A4-RSRQ and
A3-RSRP algorithms are based on the A2, A4 and A3 control events defined
G MEC Computation Handoff for Mobile Augmented Reality , UE experienced delay (ms)0.20.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (a) UE experienced delay MAD-jitter (ms)0.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (b) UE experienced MAD-jitter
UL tx throughput (Mbps)0.20.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (c) Uplink TX throughput DL rx throughput (Mbps)0.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (d) Downlink RX throughput
Figure 6: Performance improvements (x-axes on log scale) by 3GPP standard [3]. The
A2-A4-RSRQ algorithm triggers a hand-off when the serving cellโs RSRQ drops below a threshold (A2) and aneighboring cellโs RSRQ rises above an offset (A4). The
A3-RSRP al-gorithm triggers a handoff when the serving cellโs Reference SignalReceived Power (RSRP) drops below the RSRP of a neighboring cell.Both algorithms are already part of the LTE module of
NS-3 [28].The other simulation parameters are summarised in Table 3. Thedeveloped
NS-3 code is available at [4].
Improvement:
Figure 6a and Figure 6b illustrate the cumulativedistributions of UE experienced delay and jitter for all packets fromUEs. We note that since the UE experienced delay distributionsare very broad, we use a robust median-absolute deviation (MAD)version of jitter. As shown,
Comp-HO improves the UE experienceddelay with considerable improvements at the tail of the distribution.While, for jitter,
Comp-HO essentially smooths out the distribution,thus removing the large bifurcation seen in the other algorithms.
Comp-HO does this by better distributing the UEs across MECs andthus avoiding the issue of lucky UEs that happen to be in the areaswith fewer competitors for channel and MEC resources.Figure 8b and Figure 9b further support this conclusion by show-ing that with
Comp-HO the MEC servers actually process morepackets. Relatedly, the throughput performances in ?? show thatwith Comp-HO the UEs receive more processed packets (than thebenchmarks) despite actually sending slightly fewer packets to beprocessed (than the benchmarks). The UEs with Comp-HO sendslightly fewer packets packets because of a higher HO rate, whichwe discuss further in the next section.To concentrate on the majority of the performance, in the restof the paper we calculate the mean values excluding the effectof the outliers . Overall, Comp-HO outperforms the benchamarkalgorithms with improvements in mean UE experienced delay of73%-80% and downlink throughput of 3%-5%. Table 5 summarizes An outlier is more than one standard deviation from the overall mean UL SINR (mdB)0.20.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (a) Uplink SINR DL SINR (mdB)0.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (b) Downlink SINR UL delay (s)0.20.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (c) Uplink trans. delay UL STD-jitter (s)0.20.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (d) Uplink trans. STD-jitter DL delay (s)0.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (e) Downlink trans. delay DL STD-jitter (s)0.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO (f) Downlink trans. STD-jitter
Figure 7: Trade-off (all x-axes on log scale). The transmissiondelays have excluded all processing delays.Table 5: Performance improvements
Comp-HO A2-A4-RSRQ A3-RSRP NoHO
Delay (ms) 214 788 1052 1096DL TP (Mbps) 5.28 5.13 5.06 5.03UL TP (Mbps) 6.68 to 6.74 6.73 6.73MEC packets 27096 24889 24366 24433the numerical results of UE experienced delay, downlink/uplinkthroughput, and MEC processed packets.
Trade-off:
To illustrate the potential trade-off in using
Comp-HO ,Figure 7 illustrates the uplink and downlink SINR, transmissiondelay and jitter, therefore isolating transmission dynamics by re-moving processing dynamics. The results show that
Comp-HO hassomewhat lower SINR and higher transmission delay and jitter, aswould be expected given that
Comp-HO does not purely optimizeSINR. However, the increases are relatively minor in comparisonto the UE experienced delay and jitter decreases in the simulation.We also note that, as expected, the MEC-mobility packet loss of
Comp-HO is 4.3% compared to 1.2% for
A2-A4-RSRQ . This is theresult of
Comp-HO more eagerly switching base stations given highdelay MECs. However, much of this loss could be avoided in ETSIMEC networks that include MEC assisted user context transfer asthe transfer could include data (packets) waiting for processing.
Zhou, et al. U E e x p e r i e n c e d d e l a y ( m s ) Comp-HOA2A4RsrqA3RsrpNoHO (a) UE experienced delay (ms) M E C p r o c e ss e d p a c k e t s Comp-HOA2A4RsrqA3RsrpNoHO (b) MEC processed packets P a c k e t l o ss r a t i o Comp-HOA2A4RsrqA3RsrpNoHO (c) MEC-mobility packet loss ratio
Figure 8: Performance with different UE speeds (2, 6, and 10 m/s)
20 30 50Different frame rates02000400060008000 U E e x p e r i e n c e d d e l a y ( m s ) Comp-HOA2A4RsrqA3RsrpNoHO (a) UE experienced delay (ms)
20 30 50Different frame rates010000200003000040000 M E C p r o c e ss e d p a c k e t s Comp-HOA2A4RsrqA3RsrpNoHO (b) MEC processed packets
20 30 50Different frame rates0.00.10.20.30.4 P a c k e t l o ss r a t i o Comp-HOA2A4RsrqA3RsrpNoHO (c) MEC-mobility packet loss ratio
Figure 9: Performance with different frame rates (20, 30, and 50 fps)
Relatedly, the Comp-HO handoff rate is two to three times thebenchmark rates depending on the specific scenario. This causesmore service interruptions; however, when Comp-HO decides tohandoff the service is in any case usually interrupted by MECcongestion. Thus the benefit outweighs the cost. Additionally, 5Gtechniques such as Dual Active Protocol Stack (DAPS) during han-dover should minimize the length of these handoff interruptions.The other significant consideration of higher handover rates isenergy consumption, which we leave for future work.
Handoff rate:
For different effective handoff rates (as varied throughchanges in UE speed), Figure 8c shows that
Comp-HO does sufferhigher MEC-mobility packet loss ratio (PLR) comparing with otheralgorithms with slower UEs. Meanwhile, Figure 8a and Figure 8billustrate that higher handoff rates do decrease the performancegap between baseline algorithms and
Comp-HO . However,
Comp-HO still outperforms all baselines in all speeds for delay and MECload assignment performance. The decreasing gaps are actuallypartly an artifact of the increase in MEC-mobility packet loss sincemore long-delayed packets are lost and not included in the delaydistributions.
FPS:
Figure 9 illustrate the performance comparison when usingdifferent frame rates (FPS), in other words different UE sendingrates, within a common FPS range of MAR applications, i.e., 20 Hz,30 Hz, and 50 Hz, respectively. The results show that
Comp-HO improves user experienced delay significantly in all FPS, though thedegree of improvement varies likely due to an interplay of factors.Due to the focus on both MEC load and signal strength,
Comp-HO UE experienced delay (ms)0.00.20.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO
Figure 10: UE experienced delay with Gauss-Markov mobil-ity model. encounters higher packet loss ratio in higher frame rates scenario.Nevertheless, the result clearly shows that
Comp-HO becomes moresuperior in terms of MEC load balancing and UE experienced delaywhile frame rate increases, indicating the benefit and necessity of
Comp-HO for compute-intensive application offloading.
Mobility Models:
For the two different mobility models, Figures6a and 10 illustrate the UE experienced delay distributions of ran-dom waypoint and Gauss-Markov model respectively. As describedbefore, with the baseline center-biased random waypoint model(with more heterogeneous user density and MEC loads) the
Comp-HO algorithm provides significant gains. However, with the Gauss-Markov model (with more homogeneous user density and MECloads) the
Comp-HO algorithm provides very little benefit becausethere is less congestion at MEC servers and thus less need for loadredistribution.
G MEC Computation Handoff for Mobile Augmented Reality , AR task impairment0.00.20.40.60.81.0 C D F Comp-HOA2A4RsrqA3RsrpNoHO
Figure 11: AR task impairment distributions. An impair-ment score of one represents the maximum empirical per-formance while a score of zero represents the minimum,thus a higher score is better.
To illustrate the potential impact on user QoE, we utilise an existingAR task impairment model [18] to transform the delay distributionvalues into AR task impairment scores. Specifically, an impairmentscore is a normalised score that quantifies the reduction in perfor-mance (from an empirical maximum) on a task or game (e.g., gamescore), specifically in this case, collaborate assembly of a virtualAR object like in Minecraft Earth (though also with physical objectdetection in our case). In other words, an impairment score of onerepresents the maximum empirical performance while a score ofzero represents the minimum. Figure 11 illustrates the impairmentscore distributions for the different handoff algorithms. We findthat
Comp-HO significantly decreases the fraction of packets withfull impairment (indicating the lowest performance level), thussuggesting that the delay improvements from
Comp-HO shouldtranslate into actual QoE gains during these types of AR tasks.
The simulation results show that
Comp-HO significantly improvesthe user experienced delay at the expense of a small increase intransmission delay (due to a decrease in signal strength). Addi-tionally,
Comp-HO is more robust to different UE speeds and out-performs benchmark algorithms in different FPS. Comparing withhomogeneous user density and MEC loads distribution,
Comp-HO provides more improvement of user experienced delay in heteroge-neous counterpart.
In this work we proposed and evaluated
Comp-HO as a simple,effective, and standard-friendly computational handoff algorithm.In the future though, both
Comp-HO and our evaluation methodscan still be improved in several aspects.
Implementation : In terms of our evaluation, firstly, the 5G tesbeduses commercial BSs and thus due to technical and legal reasons wecould not implement
Comp-HO directly into those BSs. 5G testbedssuch as COSMOS [30] and POWDER [7] provide large scale openSDR networks, however they currently have limited access forother researchers and moreover it would be difficult to emulate UE mobility without field test. Secondly, the custom MEC-enabled
NS-3 simulator could include some features like the consideration ofuser context migration between MEC servers (e.g., the migration ofdocker containers or VMs containing the current app state [22, 44])and a more realistic MAR model at the application level.
Algorithm : In terms of improvements to
Comp-HO itself, the al-gorithm adds overheads such as collecting and sending the MECload information. Thus, in cases with very few or very spatiallyhomogeneous users, the algorithm will essentially act as a strongestSINR algorithm (since the MEC loads will be similar) but with an ad-ditional overhead and thus less efficient. Adding a threshold basedon the number or distribution of UEs could remedy this by allowingbase stations to fall back to traditional handoff algorithms duringthe mentioned conditions. Additionally, similar to having the
NS-3 simulator consider the user context migration in the evaluation,
Comp-HO could consider the user migration cost in the optimiza-tion directly (assuming differing migration costs between differentMEC pairs).
Next step : In future work, we will integrate
Comp-HO and relatedalgorithms into open Software Defined Radio (SDR) base stations,therefore allowing evaluations in a real physical deployment. Wealso plan to further develop the MEC-enabled
NS-3 simulator toallow more comprehensive simulations. We will develop algorithmsthat consider different types of interference in 5G networks andthe tradeoff between the handoff rates and the level of interferencein the network. In addition, we will develop a predictive capabilityalgorithm allowing us to predict better mobility so resources canbe pre-allocated before a handoff event even occurs.
The combination of 5G and MEC allows novel services and im-proved experiences in areas like MAR and virtual reality amongothers. Towards this goal and with a focus on mobility, this workstudied the issue of
MEC HO in 5G and proposed a handoff al-gorithm,
Comp-HO , that considers both network signal strengthand nearby MEC server load. We conducted the 5G MEC testbedmeasurements with a MAR prototype and utilized the collectedresults to set up large scale simulations with a custom MEC-enabled
NS-3 simulator. The simulation results illustrated that
Comp-HO improves the end-to-end delay compared to benchmark handoffalgorithms in MEC scenarios. As far as we know, this is one of thefirst efforts to optimize
MEC HO for 5G.
Zhou, et al.
REFERENCES [1] 3GPP. 2019.
Release description; Release 15 . Technical report (TR) 21.915. 3rdGeneration Partnership Project (3GPP). V15.0.0.[2] 3GPP. 2020.
Evolved Universal Terrestrial Radio Access (E-UTRA); Radio ResourceControl (RRC); Protocol specification . Technical specification 36.331. 3rd Genera-tion Partnership Project (3GPP). V16.1.1.[3] Evolved Universal Terrestrial Radio Access. 2013. Radio resource control (RRC).
Protocol specification (Release 10) (2013).[4] Authors. 2020. Source to be released on publication.[5] Fani Basic, Atakan Aral, and Ivona Brandic. 2019. Fuzzy handoff control in edgeoffloading. In
ICFCโ19 .[6] Christian Bettstetter, Hannes Hartenstein, and Xavier Pรฉrez-Costa. 2004. Sto-chastic properties of the random waypoint mobility model.
Wireless Networks (2004).[7] Joe Breen, Andrew Buffmire, Jonathon Duerig, Kevin Dutt, Eric Eide, Mike Hi-bler, David Johnson, Sneha Kumar Kasera, Earl Lewis, Dustin Maas, et al. 2020.POWDER: Platform for Open Wireless Data-driven Experimental Research. In
WiNTECHโ20 .[8] Dan Broyles, Abdul Jabbar, James PG Sterbenz, et al. 2010. Design and analysisof a 3โD gauss-markov mobility model for highly-dynamic airborne networks.In
ITCโ10 .[9] Derek Bruff. 2005. The assignment problem and the hungarian method.
Notesfor Math (2005).[10] P Carter and MA Beach. 1995. Evaluation of handover mechanisms in shadowedlow earth orbit land mobile satellite systems.
International Journal of SatelliteCommunications (1995).[11] Dimitris Chatzopoulos and Pan Hui. 2016. Readme: A real-time recommendationsystem for mobile augmented reality ecosystems. In
ACM MMโ16 .[12] Mostafa Zaman Chowdhury, Won Ryu, Eunjun Rhee, and Yeong Min Jang.2009. Handover between macrocell and femtocell for UMTS based networks. In
ICACTโ09 .[13] Mustafa Emara, Miltiades C Filippou, and Dario Sabella. 2018. MEC-aware cellassociation for 5G heterogeneous networks. In
WCNCโ18 .[14] European Telecommunications Standards Institute (ETSI). 2019.
Multi-accessEdge Computing (MEC); Application Mobility Service API . Recommendation 012V2.1.1.[15] Ilija Hadลพiฤ, Yoshihisa Abe, and Hans C Woithe. 2017. Edge computing in theePC: A reality check. In
SECโ17 .[16] J. Han and B. Wu. 2010. Handover in the 3GPP long term evolution (LTE) systems.In .[17] ITU-T. 2019.
Propagation data and prediction methods for the planning of short-range outdoor radio communication systems and radio local area networks in thefrequency range 300 MHz to 100 GHz . Recommendation P.1411. InternationalTelecommunication Union.[18] B. Krogfoss, J. Duran, P. Perez, and J. Bouwen. 2020. Quantifying the Value of 5Gand Edge Cloud on QoE for AR/VR. In
QoMEXโ20 .[19] Janne Kurjenniemi, Tero Henttonen, and Jorma Kaikkonen. 2008. Suitabilityof RSRQ measurement for quality based inter-frequency handover in LTE. In
ISWCSโ08 .[20] Tong Li, Chathura Sarathchandra Magurawalage, Kezhi Wang, Ke Xu, Kun Yang,and Haiyang Wang. 2017. On efficient offloading control in cloud radio accessnetwork with mobile edge computing. In
ICDCSโ17 .[21] Luyang Liu, Hongyu Li, and Marco Gruteser. 2019. Edge assisted real-time objectdetection for mobile augmented reality. In
MobiComโ19 .[22] Lele Ma, Shanhe Yi, and Qun Li. 2017. Efficient service handoff across edgeservers via docker container migration. In
SECโ17 .[23] Yuyi Mao, Jun Zhang, and Khaled B Letaief. 2017. Joint task offloading schedulingand transmit power allocation for mobile-edge computing systems. In
WCNCโ17 .[24] Marja Matinmikko, Matti Latva-Aho, Petri Ahokangas, Seppo Yrjรถlรค, and TimoKoivumรคki. 2017. Micro operators to boost local service delivery in 5G.
WirelessPersonal Communications (2017).[25] M. Mezzavilla, M. Zhang, M. Polese, R. Ford, S. Dutta, S. Rangan, and M. Zorzi.2018. End-to-End Simulation of 5G mmWave Networks.
IEEE CommunicationsSurveys & Tutorials (2018).[26] J-M Moon and D-H Cho. 2009. Efficient handoff algorithm for inbound mobilityin hierarchical macro/femto cell networks.
IEEE Communications Letters (2009).[27] Wahida Nasrin and Jiang Xie. 2019. A Joint Handoff and Offloading DecisionAlgorithm for Mobile Edge Computing (MEC). In
GLOBECOMโ19
Simulation Modelling Practice and Theory (2019).[30] Dipankar Raychaudhuri, Ivan Seskar, Gil Zussman, Thanasis Korakis, Dan Kilper,Tingjun Chen, Jakub Kolodziejski, Michael Sherman, Zoran Kostic, XiaoxiongGu, et al. 2020. Challenge: COSMOS: A city-scale programmable testbed forexperimentation with advanced wireless. In
MobiComโ20 . [31] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. 2016. You Only Look Once:Unified, Real-Time Object Detection. In
CVPRโ16 .[32] Debashis Saha, Amitava Mukherjee, Iti Saha Misra, and Mohuya Chakraborty.2004. Mobility support in IP: a survey of related protocols.
IEEE network (2004).[33] Stefania Sardellitti, Gesualdo Scutari, and Sergio Barbarossa. 2015. Joint optimiza-tion of radio and computational resources for multicell mobile-edge computing.
IEEE Transactions on Signal and Information Processing over Networks (2015).[34] M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies. 2009. The Case forVM-Based Cloudlets in Mobile Computing.
IEEE Pervasive Computing (2009).[35] Wei Shen and Qing-An Zeng. 2006. A novel decision strategy of vertical handoffin overlay wireless networks. In
NCAโ06 .[36] Kari Sipila, Mika Jasberg, Jaana Laiho-Steffens, and Achim Wacker. 1999. Softhandover gains in a fast power controlled WCDMA uplink. In
VTCโ99 .[37] Nishith D Tripathi, Jeffrey H Reed, and Hugh F VanLandinoham. 1998. Handoffin cellular systems.
IEEE personal communications (1998).[38] Zi Wang, Zhiwei Zhao, Geyong Min, Xinyuan Huang, Qiang Ni, and Rong Wang.2018. User mobility aware task assignment for mobile edge computing.
FutureGeneration Computer Systems (2018).[39] Peng Xu, Xuming Fang, Rong He, and Zheng Xiang. 2013. An efficient handoffalgorithm based on received signal strength and wireless transmission loss inhierarchical cell networks.
Telecommunication Systems (2013).[40] Peng Xu, Xuming Fang, Jun Yang, and Yaping Cui. 2010. A userโs state andSINR-based handoff algorithm in hierarchical cell networks. In
WiCOMโ10 .[41] Candy Yiu, Yujian Zhang, Mo-Han Fong, and Yuefeng Peng. 2015. User equipmentand methods for handover enhancement using reference signal received quality(RSRQ). US Patent 9,130,688.[42] Fangxiaoqi Yu, Haopeng Chen, and Jinqing Xu. 2018. DMPO: Dynamic mobility-aware partial offloading in mobile edge computing.
Future Generation ComputerSystems (2018).[43] Zikang Yuan, Dongfu Zhu, Cheng Chi, Jinhui Tang, Chunyuan Liao, and XinYang. 2019. Visual-Inertial State Estimation with Pre-integration Correction forRobust Mobile Augmented Reality. In
ACM MMโ19 .[44] Aleksandr Zavodovski, Nitinder Mohan, Suzan Bayhan, Walter Wong, and JussiKangasharju. 2018. Icon: Intelligent container overlays. In
HotNetsโ18 .[45] Wenxiao Zhang, Bo Han, and Pan Hui. 2018. Jaguar: Low latency mobile aug-mented reality with flexible tracking. In
ACM MMโ18 .[46] Ziyue Zhang, Jie Gong, Xiang Chen, and Terng-Yin Hsu. 2020. ReinforcementLearning Based Computation-aware Mobility Management in Ultra Dense Net-works.