Mobility Management for TCP in mmwave Networks

Size: px

Start display at page:

Download "Mobility Management for TCP in mmwave Networks"

Lilian Watson
5 years ago
Views:

1 Mobility Management for TCP in mmwave Networks Michele Polese *, Marco Mezzavilla v, Sundeep Rangan v, Michele Zorzi * * Dept. of Information Engineering, University of Padova, Italy V NYU Wireless, Brooklyn, NY, USA michele@polese.io October 16, 2017

2 Outline Introduction TCP in mmwave cellular networks Mobility management architectures Performance evaluation ns-3 mmwave module Results Conclusions

3 mmwave cellular networks Part of 3GPP New Radio PHY-layer issues impact the higher layers Small cells Beamforming Blockage Joint performance analysis of transport layer and mobility in mmwave cellular networks

4 TCP in mmwave cellular networks TCP most used transport protocol (so far..) Loss-based congestion control Performance on wireless networks has been investigated since the 90s mmwave cellular networks introduce new challenges

5 Challenges for TCP in mmwave Very high bandwidth Issues with congestion window slow ramp-up Extended outages Retransmission timeouts and resets LOS/NLOS link variability Bufferbloat Cross-layer approaches? Multipath TCP? Rely on smart network management?

6 Requirements for TCP and mobility Prompt reaction to channel updates Continuous coverage Availability of multiple beams Minimize Packet loss Handover interruption time Low end-to-end latency

7 Mobility management in mmwave Stand-alone Single connectivity Traditional Hard Handover (HH)

Mobility management in mmwave Dual-connectivity LTE overlay + mmwave base station Fast switch + faster secondary cell handover M. Polese, M. Giordani, M. Mezzavilla, S.

8 Mobility management in mmwave Dual-connectivity LTE overlay + mmwave base station Fast switch + faster secondary cell handover M. Polese, M. Giordani, M. Mezzavilla, S. Rangan and M. Zorzi, "Improved Handover Through Dual Connectivity in 5G mmwave Mobile Networks," in IEEE Journal on Selected Areas in Communications, vol. 35, no. 9, pp , Sept. 2017

Connectivity Different server deployment scenarios Core network /

9 Performance evaluation Comparison of Single base station scenario (no handover) Single Connectivity with Hard Handover Dual Connectivity Different server deployment scenarios Core network / Internet D = 10 / 20 ms Remote server Mobile Edge Cloud server D = 1 ms

TCP/IP RRC PDCP RLC MAC PHY CHANNEL MODEL BEAMFORMING Base station

10 ns-3 mmwave module Based on ns-3 + LTE module End-to-end performance analysis 3GPP mmwave channel implementation Base station UE APP TCP/IP RRC PDCP RLC MAC PHY CHANNEL MODEL BEAMFORMING Base station Tunneling RRC PDCP RLC MAC PHY CN function MME PGW/SGW TCP/IP APP Remote Server

11 Laboratorio di Fondamenti di Informatica Y[ (anduethus high data rate) available at the physical layer (incnthe function Base station 60 order of Gbit/s), but the currently available TCP congestion APP MME 40 TCP/IP algorithms in high BDP scenariostunneling control o er a decreasing PGW/SGW CHANNEL MODEL throughput as the end-to-end latency increases [19]. 20 RRC RRC BEAMFORMING UE path at PDCP PDCP UE In this paper, we consider three di erent mobility managespeed RLC RLC 0 ment MACschemes. The baseline is a basic strategy, MAC in which the TCP/IP APP UEPHY connects to the mmwave base station PHY with the highest X [m] Signal to Noise Ratio (SNR) and as it moves it does not update Remote Server Figure 2: Example of simulation scenario. The grey rectangles are 15 the serving access point (no handover). We look at the no hanrandomly deployed non-overlapping obstacles (e.g., cars, buildings, dover case to demonstrate the value of dense deployments people, trees). Figureand 1: End-to-end protocol stackthis considered the performance macro diversity. Moreover, baseline isin equivalent evaluation. Parameter Value to the deployment considered in [22]. We then consider a single connectivity approach, in which each mmwave base mmwave carrier frequency 28 GHz station is directly connected to the core network, and a HH mmwave bandwidth 1 GHz 120 to update the serving mmwave + LTE base station or is required mmwave LTE carrier frequency (DL) 2.1 GHz mmwave mmwave base stations fall back to the legacy LTE RAT (Single Connectivity with base station base station LTE bandwidth 20 MHz 100 3GPP Channel Scenario Urban Micro HH ). The last scheme is the dual connectivity architecture mmwave outage threshold 5 db proposed 80 in [11] (Dual Connectivity (DC)). Y [m] Scenario 60 4 PERFORMANCE EVALUATION 4.1 Simulator and Scenarios 40 mmwave max PHY rate X2 link latency D X 2 S1 link latency D S 1 PGW to remote server latency D RS RLC bu er size B RLC RLC AM reordering timer S1-MME link latency D M M E UE speed Number of obstacles N obs TCP Maximum Segment Size 3.2 Gbit/s 1 ms 1 ms [0, 10, 20] ms 1 MB 1 ms 10 ms 5 m/s [5, 15] 1400 byte For the performance evaluation we run a simulation cam20 paign using the end-to-end mmwave module for UE ns 3 path atdeue speed scribed in 0 [6] with the dual connectivity extension presented in [13]. The0 simulated 50 3GPP-like protocol stack and the dif ferent nodes are presented in Fig. 1. An example of scenario X [m] is shown in Fig. 2. There are three mmwave and one LTE Table 1: Simulation parameters base stations, and N obstacles of di erent in rectangles the area Figure 2: Example of simulation scenario. Thesize grey are 15 obs between the base stations and the user. The obstacles arebuildings, one proposed by 3GPP in [1], with the spatial consistency randomly deployed non-overlapping obstacles (e.g., cars, placed in the scenario randomly in each simulation run, in oroption, so that as the user moves the channel matrix is uppeople, trees). der to capture di erent possible propagation environments. dated in a correlated way. The user moves at speed along Randomly generated in people eachand run or 15 the obstacles) They model buildings, trees, or other they(5 force horizontal path from x = 50 m to x = 150 m, and then Parameter Value a NLOS condition when interposed between the user and a turns back and repeats the path multiple times, so that it is base station. Moreover, the link is considered to28 beghz in outage possible to measure the performance of TCP in steady-state. mmwave carrier frequency if the SNR is below a threshold. The channel model is the The application layer simulates a le transfer with full bu er. mmwave bandwidth 1 GHz

12 Goodput 0 Mean Mean First quartile First quartile Third quartile Third quartile One-way end-to-end latencylatency DS 1 + D D SRS One-way end-to-end D RS [ms] 1 +[ms] Goodput [Mbit/s] Goodput [Mbit/s] Goodput [Mbit/s] 500 1,500 1,500 Dual Connectivity Dual Connectivity No handover No handover 1,000 1,000 Goodput [Mbit/s] 1,500 1,500 Mean Mean First quartile First quarti Third quartile Third quart 1,000 1, Dual Connectivity Dual Connectivity Single Connectivity (HH) (H Single Connectivity One-way end-to-end latencylatency DS 1 + D D SRS One-way end-to-end D RS [ms] 1 +[ms] FigureFigure 3: Goodput for a scenario with Nwith random obstacles, obs =N15 3: Goodput for a scenario obs = 15 random obstacles, with and without handovers. with and without handovers. 15 obstacles 2,000 2,000 Goodput [Mbit/s] Goodput [Mbit/s] We compare three three di erent valuesvalues for the D RS of We compare di erent forlatency the latency D RS of the wired link between the Packet Gateway (PGW) and the the wired link between the Packet Gateway (PGW) and the application server, in both a Mobile Edge Cloud (MEC) sceapplication server, in both a Mobile Edge Cloud (MEC) scenario nario with an Edge ServerServer (ES) (i.e., 0 ms the the RS = with an Edge (ES)D(i.e., D RS = since 0 ms since serverserver is deployed in theincore and a and scenario with with is deployed thenetwork) core network) a scenario a Remote ServerServer (RS) (i.e., 20 or ms). RS = a Remote (RS)D(i.e., D 10 = 10 20HighSpeed ms). HighSpeed RS or TCP is used control, since since it wasitdesigned TCP is for usedcongestion for congestion control, was designed for high product networks [5]. The for bandwidth-delay high bandwidth-delay product networks [5].main The main (a) N obs(a)=n15obsrandom obstacles. = 15 random obstacles. Dual Connectivity Dual Connectivity Single Connectivity (HH) (HH Single Connectivity Mean Mean First quartile First quartile Third quartile Third quarti Dual and single connectivity -> better than no handover 1,000 (edge server) Impact of end-to-end latency 1,000

13 RAN latency [s] 15 obstacles Median RAN latency [s] vity 4.3 Latency more Fig. 5 reports the boxplots for the RAN latency of successfully obreceived packets at the PDCP layer, for di erent mobility xed 10 2 management schemes and di erent values of N obs. It can be vity immediately seen that adapting the serving base station to ngle 2.5 the best10one available not only increases the goodput, but also reduces the latency. The handover procedures may occae in sionally introducenoadditional because(hh) of thedual handover Handover latency Connectivity Single Connectivity with RAN latency [s] Latency No Handover Single Connectivity (HH) Dual Connectivity (a) N obs = 15 random obstacles. 10 RAN latency [s] on average for the entire runs where handover events are relatively infrequent. Thus, the di erence in average throughput is not large. We will see in the next section that the more signi cant gain is in latency. In general, the dual connectivity option manages to complete the handovers between mmwave base stations or the switches across RATs in a shorter time, with fewer packet losses, therefore it sustains a generally higher goodput. However, the single connectivity solution manages to reach a better performance when there is a short interval of time with the channel in LOS condition base station. In vely and the user does not change the serving Mean of the single connecured this case, indeed, the overall latencymedian is smaller than that of the dual connectivity are tivity option ugh- deployment, therefore the congestion window grows more more quickly. In the scenario with N obs = 5 and the ES, we obnec- served that, if the same latency is considered in the xed network, then the solution with dual connectivity een part of the 10 2 in a gains on average 400 Mbit/s (20%) with respect to the single ns a connectivity architecture. of obstacles N obs plays a major role in vity Finally, the number No Handover Single Connectivity (HH) Dual Connectivity here the achievable goodput, which is up to 2 times higher with 15.Nobs In =the case, indeed, there is a 15 rst random obstacles. tion 5 obstacles than with(a) of having a LOS channel, thus a higher n. In higher probability 1 Mean 10 nec- data rate available at the physical layer Mean Median No Handover Single Connectivity (HH) Dual Connectivity (b) N obs = 5 random obstacles. 5 obstacles Figure 5: RAN one-way latency for the three di erent mobility management schemes, with a di erent number of obstacles N obs. Notice that the y-axis is in logarithmic scale. Finally, if we consider the two architectures in which the handovers are allowed, the one with dual connectivity manages to keep the latency at a minimum, and with a smaller variability as shown by the boxplots, thanks to the faster handover or RAT switch procedures [11]. No handover -> bufferbloat Dual connectivity (fast handovers no service interruption) -> lowest RAN latency 4.4 RLC AM and RLC UM

14 Edge server: RLC AM or UM? goodput latency DC with RLC AM -> highest goodput and smallest latency

15 Conclusions End-to-end evaluation of TCP, mmwave, mobility Multiple base stations + fast handover procedures improve both goodput and latency No bufferbloat! Edge server gives the best goodput performance Dual connectivity allows to reduce latency Next steps: TCP proxy -> improve TCP reactiveness Cross-layer approaches Real testbed

16 Useful resources ns-3 mmwave module mmwave (branch new-handover for DC) mmwave cellular + vehicular UNIPD NYU Wireless

17 Mobility Management for TCP in mmwave Networks Michele Polese *, Marco Mezzavilla v, Sundeep Rangan v, Michele Zorzi * * Dept. of Information Engineering, University of Padova, Italy V NYU Wireless, Brooklyn, NY, USA michele@polese.io October 16, 2017

X-TCP: a Cross Layer Approach for TCP Uplink Flows in mmwave Networks Laboratorio di Fondamenti di Informatica

X-TCP: a Cross Layer Approach for TCP Uplink Flows in mmwave Networks Laboratorio di Fondamenti di Informatica Tommy Azzino, Matteo Drago, Michele Polese, Andrea Zanella, Michele Zorzi June 28, 2017 Dept.