12 12 Italy Padova, of University Zorzi Michele AT&T Labs–Research Rittwik Jana Italy Padova, of University PoleseMichele mmWave Networks TCP and MP-TCP in 5G Society Computer IEEE the by Published target of 5G networks. At frequencies frequencies At of networks. 5G target throughput the meeting in a major role (MTC). tions communica machine-type massive for low-power support and nications; of commu availability and reliability 10 below ms); high (possibly end-to-end ultra-low conditions; able favor most the in ond (Gbit/s) peaks sec per gigabits with edges, at cell (Mbit/s) second per megabits 50 at least provide the following the provide shall networks 5G alliance, (NGMN) Network Mobile Next-Generation the to According growth. traffic’s T Long-Term as such links, (LTE) various across Evolution mmWave. and TCP (MP-TCP) Multipath judiciously when used is tradeoff throughput-latency TCP, protocol, the on mmWave and transport insights provides also It links. used widely most the between interaction by asuboptimal affected be could mmWave mobile in experience user networks end-to-end how the explains article This quality. signal the received in fluctuations wide to lead which conditions, channel dynamic by highly are characterized frequencies these However, capacity. multi-gigabit-per-second potential their of because links, mmWave include likely will communication access 5GFuture radio networks MmWave communications will play play will MmWave communications before 2020, to address mobile mobile to address 2020, before standardized be (5G) will works net of mobile generation next he 1 ©2017 IEEE 1089-7801/17/$33.00 : a user bitrate of bitrate : auser - - - - - body human the even and cars, buildings, —for example, materials solid most by blockage and loss path isotropic high from suffer frequencies these fact, In market-ready. technology this to make issues that researchers must address and challenges numerous present mmWaves However, networks. lular to cel allocated be can that spectrum of contiguous availability ahigh is there indeed, (GHz), 10above gigahertz to provide a reliable service in the the in service areliable to provide how but loss; path for the up make can (MIMO), output multiple input, multiple massive and beamforming as such techniques, antenna multiple using by provided gain antenna high The outage). an is, (that unavailability 2 — which could result in service service in result could —which IEEE INTERNET COMPUTING INTERNET IEEE - TCP and MP-TCP in 5G mmWave Networks

presence of frequent blockages is still an open control has had several evolutions: question. In particular, a sudden transition the Request for Comments (RFC) 56814 describes of a mmWave link from a line-of-sight (LOS) the most recent one, and there are 13 TCP vari- to a non-line-of-sight (NLOS) channel state ants currently implemented in the ker- generates wide fluctuations in the signal-to- nel.5 In particular, the latest versions of the interference-plus-noise ratio (SINR; in the order major operating systems use TCP CUBIC as the of 30 dB, according to Sundeep Rangan and default.6 This TCP version was designed in order colleagues2), and therefore the capacity offered to react more promptly to packet losses, and to changes drastically, thereby compromising the restore the connection to the highest available end-to-end user experience. rate in a shorter time than legacy designs based These extreme propagation conditions on TCP New Reno. In the latter, after a packet demand a new design of the physical and loss, the sender decreases its congestion win- medium access control (MAC) layers, but also dow by half, and then increases it by one packet have an impact on the interplay with the in each round-trip time (RTT), thus requiring a higher layers of the protocol stack. The most very long time to fill a high capacity link. With widely used reliable data transport protocol TCP CUBIC, instead, the growth is independent is TCP, which however was designed in a dif- of the RTT, which makes it possible to reach a ferent historical context, and according to higher throughput more quickly than in New different needs and constraints. TCP considers Reno, while behaving fairly toward flows using packet losses as an implicit notification of con- other versions of TCP.6 gestion on the link, and therefore reduces the Another trend in modern - sending rate trying to relieve it. However, the ing is to exploit the presence of multiple net- lossy nature of mmWave links might trigger work interfaces in mobile devices. A typical the TCP congestion control mechanisms even if example is the usage of MP-TCP for vertical and there is no actual congestion. This might yield a seamless handovers between third- and fourth- suboptimal end-to-end performance and waste generation (3G/4G) cellular networks and Wi-Fi the great potential of mmWave links. hotspots.7 MP-TCP is an extension of TCP that With this in mind, here we provide an over- enables multipath transport — that is, the appli- view of the impact of mmWave communications cation communicates with a traditional TCP on the performance of TCP, outlining which are socket, which transparently handles multiple the main novelties that the usage of a mmWave subflows on different interfaces, such as Wi-Fi, radio access network (RAN) introduces in end- the cellular network, and Ethernet. MP-TCP is to-end connections at the . We currently under discussion in the IETF,8,9 and also consider how adopting the recent Multipath is designed around three main goals.10 First, TCP (MP-TCP) option can improve the perfor- it should improve the throughput, in the sense mance of TCP on these kinds of links. Finally, that it should perform at least as well as a single we suggest possible future research directions. path TCP (SP-TCP) flow on the best path avail- A more technical discussion with additional able. Second, it should not use more resources results is provided elsewhere.3 than standard TCP flows. Finally, it should steer more packets towards less congested paths. Recent Advances in TCP The core of MP-TCP is its congestion control TCP was designed in the 1980s as a connection- , that manages the congestion win- oriented and reliable protocol that provides dow of each different subflow in a coupled or end-to-end connectivity over multiple hops and uncoupled manner. With an uncoupled conges- congestion control (CC). Reliability is enabled tion control, each subflow is independent, thus by a retransmission mechanism, based on the the congestion window is updated separately in acknowledgments received by the TCP transmit- each path. With a coupled approach, instead, ter. Congestion control is implemented by dif- the congestion window of the different subflows ferent algorithms that increase and/or decrease is increased and decreased in a coordinated the maximum amount of unacknowledged data manner, considering the congestion of all the that the sender is allowed to transmit (the con- available subflows. By coupling the different gestion window), reacting to network events subflows, Costin Raiciu and colleagues10 claim such as packet losses. The original congestion that it is possible to reach the aforementioned

SEPTEMBER/OCTOBER 2017 13 5G

networks with mmWave links, showing how 50 mmWave 28 GHz state-of-the-art transport protocols interact LTE 2.1 GHz with this technology. 40 The mmWave Channel 30 Because TCP is the most widely used trans- port protocol, it is important to understand the (dB) 20 interactions that exist between mobile networks

SIN R (for example, wireless channels) and the TCP 10 performance. In wireless networks, the loss of a packet is not caused necessarily by congestion, 0 but instead might be due to a sudden (and pos- sibly only temporary) drop in signal quality. In 13,14 Ϫ10 related work, the authors study the behavior 0.5 1 1.5 2 2.5 3 3.5 4 4.5 Time (s) of TCP in relation to a complex mobile network such as LTE, showing the following: first, as the Figure 1. The ns-3 simulated signal-to-interference-plus-noise ratio distance between the user equipment (UE) and (SINR) for a mmWave and Long-Term Evolution (LTE) link, with the evolved node base (eNB) increases, the TCP the user equipment moving at 2 meters per second (m/s) from throughput degrades; and second, how TCP is (45, 0) to (55, 0), with the evolved node base at coordinates affected by network events such as handovers. (75, 50). The traces are generated using the channel model for the MmWave networks are expected to reach mmWave frequencies (described elsewhere2,15) and the ns-3 LTE an order of magnitude higher throughput than channel model. current systems, thanks to the larger available, but present more troublesome prop- agation conditions. As an example, Figure 1 shows a comparison between the time evolution MP-TCP design goals. The first coupled con- of the SINRs of a mmWave link at 28 GHz and gestion control (CC) that they proposed,10 how- an LTE link at 2.1 GHz for a UE that moves at ever, received criticism,11,12 because it transmits 2 meters per second (m/s) at an average distance too much traffic on congested paths and is of 75 meters from the eNB and switches from a unfriendly with respect to SP-TCP. Ramin LOS to a NLOS condition. The main differences Khalili and colleagues11 presented the Oppor- between the mmWave and the LTE channels are tunistic Linked Increases Algorithm (OLIA), as follows: which is designed to overcome these two issues. However, according to Qiuyu Peng and col- • The LOS to NLOS path loss transitions are leagues,12 OLIA presents unresponsiveness deeper for mmWave. Below 6-GHz frequen- problems with respect to congestion changes in cies, these are of the order of 10–15 dB, the subflows. The Balanced Linked Adaptation while for the mmWave channel the fluc- algorithm (BALIA)12 addresses both the prob- tuation can exceed 30 dB. Therefore, the lems of the original CC and those of OLIA. In available capacity changes dramatically. particular, the parameters of the protocol are Moreover, mmWave networks will be small derived through a theoretical analysis of the cell networks, and mmWave links are sen- performance of multipath congestion control sitive to blockage from foliage, the human algorithms. However, these schemes are based body, moving obstacles and so on, thus the on the legacy design of Reno and New Reno transitions from LOS to NLOS for mmWave congestion control algorithms, and thus share connections will be much more frequent with them the drawbacks we described in the than in LTE.2 Furthermore, given the shorter previous paragraph. range of mmWave communication, it is more probable for mmWave links than for LTE TCP Performance on ones to experience an outage (where no sig- mmWave Networks nal is received) because of shadowing. In this article, we discuss the main issues that • The mmWave channel has a shorter coher- arise when TCP and MP-TCP are deployed in ence time, which results in variations of the

14 www.computer.org/internet/ IEEE INTERNET COMPUTING TCP and MP-TCP in 5G mmWave Networks

channel on the order of hundreds of micro- throughput. When using mmWave links, these seconds,2 that are faster than those in cur- retransmission protocols become a key element rent mobile networks. As Figure 1 shows, the in hiding the highly dynamic and consequently transitions of the LTE channel due to fad- unstable behavior of the channel to higher- ing are much smoother than those of the layer transport protocols such as TCP. mmWave link. The MAC layer uses a Hybrid Automatic Repeat reQuest (HARQ) mechanism. When the Currently, these issues are mainly addressed physical layer at the receiver receives a packet, in the RAN, but they also impact the perfor- but detects the presence of some errors that mance of transport layer protocols. To date, prevent reliable decoding, it asks for a retrans- there are only few published results on the per- mission. The sender then transmits additional formance of TCP over 5G mmWave links. There redundancy that helps retrieve the packet with- are several reasons for this. The first is the lack out any error.18 Morever, in 3GPP-like networks of mobile testbeds that scale beyond a single (such as LTE), there is a layer on top of the MAC mmWave link. There are indeed commercial layer that might perform additional link-level prototypes that can be used to set up a mmWave retransmissions — the radio link control (RLC) link between a MIMO transmitter and receiver, layer — that will also likely be a part of the 5G but the focus of these products is on the single protocol stack.19 Because the number of retrans- connection. The deployment of large testbeds is missions at the MAC layer is usually limited limited by the lack of a definitive standard and (typically only three attempts are performed), by the cost and complexity of prototyping the the RLC layer acknowledged mode (AM) offers electronics for these links. Moreover, an end- another way of recovering lost packets. Thanks to-end system composed of a mmWave RAN, a to periodic acknowledgments from the receiver, core network (CN), and the Internet with remote the RLC AM sender knows which packets were servers has an extremely high complexity, and lost and can retransmit them. The number this prevents an insightful theoretical analysis of attempts that RLC AM can perform is also of the system. Finally, an analytical channel limited. The RLC unacknowledged mode (UM) model that represents the correlation of fading instead does not perform any retransmission over time in mmWave links has not been devel- in addition to those of the HARQ at the MAC oped yet. layer. Both HARQ and RLC AM operate based As mmWave networks and 5G are not on information related to the link, and with a currently standardized, a good compromise greater timeliness with respect to TCP, which between the realism of a testbed and the conve- uses packet losses to detect congestion and nience of a theoretical analysis is system-level operates on the larger timescale of retransmis- simulation, which can provide deep insights on sion time-outs (RTOs), of the order of a second. a complex system while avoiding the need for Menglei Zhang and colleagues presented expensive dedicated hardware. In this article, a first study of TCP in mmWave cellular net- we use an extension16 of the NYU mmWave17 works.20 The authors studied different sce- module of ns-3, a widely used open source narios, with either a statistical channel model, network simulator, together with the TCP/IP based on real measurements in New York City, stack of the Linux kernel, to provide a realis- or a raytracing model, with traces from a route tic insight on the interplay between TCP and in Bristol. They observed that, in the pres- mmWave links. ence of long outage events, neither HARQ nor RLC AM retransmissions are able to mask the Link-Level Retransmissions and TCP channel losses to the TCP sender, which trig- One of the main issues of mmWave links is the gers an RTO event, after which it might take a highly dynamic channel that, because of block- long time to recover the full capacity. Moreover, ages, experiences wide fluctuations of the per- because of the presence of these retransmission ceived SINR and therefore is characterized by mechanisms, TCP does not promptly track the unreliable transmissions. Current and future mmWave link state when the channel between mobile networks deploy different retransmis- the UE and the eNB switches from an LOS to sion mechanisms to mitigate the effects of NLOS condition. This issue is more marked in and increase the mobile devices’ mmWave than in current cellular networks,

SEPTEMBER/OCTOBER 2017 15 5G

mmWave UM, no HARQ Latency mmWave UM Throughput 250 3,000 mmWave AM HARQ ϩ RLC AM 200 HARQ ϩ RLC UM 0.5 No HARQ ϩ RLC UM (ms) 2,000 y 150 laten c Download time (s) 100 1,000 0 Throughput (Mbit/s) End-to-en d 50 10 140 5 120 0 0 100 80 50 75 100 150 60 File size (Mbyte) 1 a Distance (m) Distance (m) Figure 2. TCP performance on mmWave links, with and without link-level retransmissions. (a) TCP latency (dotted bars) and throughput, with and without the different retransmission mechanisms of the mmWave protocol stack. (b) Download time as a function of the file size and of the distance, for TCP with and without lower-layer retransmissions. (AM stands for acknowledged mode; HARQ stands for Hybrid Automatic Repeat reQuest; RLC stands for radio link control; and UM stands for unacknowledged mode.)

because LOS to NLOS transitions are much retransmissions are combined at very large more frequent and provide wider variations in distances. the link capacity.2 The consequence is that the If we consider the download of small files packets queue in the RAN buffers, and latency (from 1 to 10 Mbytes) using wget, which can increases. This is an instance of the well-known be representative of web browsing, then, as problem21: because of the large size Figure 2b shows, the difference between the of buffers in network devices, packets are not download times with RLC AM (both RLC dropped but queued and eventually transmit- retransmissions and HARQ) and with RLC UM ted. While this improves the throughput of (HARQ only) is more noticeable than that the mmWave link, it also increases the overall between the values of Figure 2a, showing that, latency. for short-lived TCP sessions, it is important to This tradeoff can be better understood by perform retransmissions as quickly as possible — considering the results in Figure 2a, where we that is, at a layer as close to the radio link as use the iPerf network bandwidth measurement possible. tool on top of TCP CUBIC with and without The design goals of 5G networks, nonethe- link-level retransmissions, at a different dis- less, demand both high throughput and low tance d between the UE and the eNB. At a small latency. However, as shown in Figure 2a and distance (that is, d 5 50 m), the achievable other work,20 when the distance between eNB TCP throughput is high even without retrans- and UE is high or there are blockages and an missions, because the UE is in LOS with very NLOS channel, a high TCP throughput can high probability. As the distance increases, be reached only with retransmissions, thus two different behaviors can be observed, increasing latency. The latency-throughput if retransmissions are used: the achievable tradeoff is an issue that must be solved before throughput is from 1.5 to 3 times higher than mmWave networks can be considered market- without retransmissions; however, the latency ready. A first step, proposed in related work,22,23 increases, especially when MAC and RLC layer is a cross-layer approach that makes TCP aware

16 www.computer.org/internet/ IEEE INTERNET COMPUTING TCP and MP-TCP in 5G mmWave Networks

Latency Throughput 600 4,000 LTE ϩ 28 GHz mmWave, CUBIC

500 LTE ϩ 28 GHz mmWave, BALIA 28 GHz ϩ 73 GHz mmWave, CUBIC 3,000 28 GHz ϩ 73 GHz mmWave, BALIA 400 (ms) 28 GHz mmWave, CUBIC y (Mbit/s)

laten c 300 2,000

200 Throughpu t End-to-en d 1,000 100

0 0 50 100 150 Distance (m)

Figure 3. Multipath TCP (MP-TCP) throughput and latency for different distances d and different MP-TCP options. The rightmost bar in each group shows the performance of a single path TCP (SP-TCP) connection with TCP CUBIC over a 28-GHz mmWave link as a reference. (BALIA stands for Balanced Linked Adaptation algorithm.)

of the actual channel conditions. A second throughput performance of different MP- solution might involve the replacement of TCP TCP congestion control algorithms over with other transport protocols, or the redesign different connections, with respect to the of HARQ and RLC-layer retransmissions in the baseline of a SP-TCP connection with TCP 3GPP stack. Finally, an ultra-dense deployment CUBIC on a mmWave link. When the UE has could decrease the probability of having NLOS a high probability of being in LOS (that is, links toward mmWave eNBs; however, in this for d # 50 m), the solution with MP-TCP on case the issue becomes a robust management of mmWave-only links outperforms SP-TCP, mobility, handovers, and cell association.15 with a gain between 800 Mbit/s and 1 Gbit/s (about 30–40 percent). The LTE link, instead, Multipath TCP in mmWave Networks has a much smaller rate than a mmWave link 5G devices will likely be connected to multiple when d # 50 m, and therefore the through- Radio Access Technologies (RATs) or to mul- put of MP-TCP on LTE and mmWave subflows tiple mmWave eNBs.15 This is why it is inter- is close to or worse than that of the reference esting to analyze the performance of MP-TCP SP-TCP. However, the performance of an LTE over mmWave and LTE links, since it could link is less dependent on the distance d than be used as an end-to-end solution for multi- that of a mmWave link; thus, for higher dis- connectivity. However, the complex interac- tances, MP-TCP over LTE and mmWave links tions that exist between the reliability of the performs better than MP-TCP with only different paths and the different congestion mmWave connections. The 73-GHz mmWave control algorithms require a deep understand- subflow, indeed, provides a larger theoreti- ing of MP-TCP performance before deploying cal capacity than the LTE one, but has a lossy it on mmWave links. behavior that penalizes the overall throughput, First, we consider whether using LTE or except for small distances. As Figure 3 shows, mmWave as a secondary subflow yields a for d 5 150 m, MP-TCP with LTE and 28 GHz higher throughput. In Figure 3 we show the mmWave offers a gain of more than 450 Mbit/s

SEPTEMBER/OCTOBER 2017 17 5G

(about 100 percent) with respect to SP-TCP An optimized interaction between trans- (that is, more than the LTE link throughput), port protocols and mmWave networks will without increasing the latency. This shows that be a fundamental enabler of 5G networks. In link diversity —that is, the presence of the sec- particular, the main issue is related to how to ondary and reliable LTE path — improves the reduce latency while maintaining a high and overall connection’s throughput. stable throughput while using TCP and MP- Second, we compare the performance of two TCP over mmWave networks, so that the end different congestion control algorithms for MP- users can effectively exploit the huge amount TCP: BALIA (that is, a coupled CC algorithm), of resources available in fifth-generation cel- and CUBIC (an uncoupled algorithm). As Figure lular networks. 3 shows, MP-TCP with the coupled BALIA CC To reach this goal, in the next few years algorithm fails to meet the first target of MP-TCP the following open research questions must be design, because in many cases its throughput addressed. How do we design reliable hando- is lower than that of SP-TCP. When coupling a ver mechanisms that do not impact TCP per- mmWave and an LTE link, indeed, the congestion formance in high mobility scenarios? How control algorithm sees the losses on the 28-GHz can we optimally exploit multi-connectivity, mmWave connection as congestion, and steers either with MP-TCP, or with different solutions the whole traffic to the LTE subflow, degrad- at a lower layer in the RAN, also studying the ing the end-to-end connection’s performance. trade­off between energy consumption and very Instead, the uncoupled congestion control algo- high throughput? And how can we combine the rithm is unaffected by this issue, because each stability and reactiveness of TCP CUBIC with a path behaves independently. However, in this MP-TCP congestion control algorithm coupled case the MP-TCP flow might be unfriendly to over 5G mmWave and legacy links? SP-TCP flows on shared bottlenecks. Therefore, the definition of a CC algorithm References that meets the three design goals of MP-TCP for 1. M. Iwamura, “NGMN View on 5G Architecture,” Proc. mmWave links is still an open area of research. 2015 IEEE 81st Vehicular Technology Conf., May 2015, The current CC algorithms either waste the poten- pp. 1–5. tial gain given by multi-connectivity, because of 2. S. Rangan, T.S. Rappaport, and E. Erkip, “Millimeter- suboptimal balancing between flows, or harm the Wave Cellular Wireless Networks: Potentials and Chal- fairness among network users. Therefore, before lenges,” Proc. IEEE, vol. 102, no. 3, 2014, pp. 366–385. deploying these algorithms in future mmWave 3. M. Polese, R. Jana, and M. Zorzi, “TCP in 5G mmWave networks, a proper congestion control algorithm Networks: Link Level Retransmissions and MP-TCP,” should be designed and thoroughly tested. Proc. 2017 IEEE Conf. Computer Comm. Workshops, 2017; https://arxiv.org/pdf/1703.08985.pdf. 4. M. Allman, V. Paxson, and E. Blanton, TCP Congestion e provided an overview of the issues that Control, IETF RFC 5681, 2009; https://tools.ietf.org/ W might arise from the complex interactions html/rfc5681. between TCP — either single or multipath — and 5. C. Callegari et al., “A Survey of Congestion Control mmWave links. The interplay between commu- Mechanisms in Linux TCP,” Distributed Computer and nications at such high frequencies and legacy Comm. Networks, Mar. 2014, pp. 28–42. transport protocols could be a limitation that 6. S. Ha, I. Rhee, and L. Xu, “CUBIC: A New TCP-Friendly prevents the full exploitation of the potential High-Speed TCP Variant,” ACM Sigops Operating Sys- of mmWave communications. For mmWave, it tems Rev., vol. 42, no. 5, 2008, pp. 64–74. is fundamental to mask the channel losses to 7. Y.-C. Chen et al., “A Measurement-Based Study of Mul- the higher TCP layer with link-level retrans- tipath TCP Performance over Wireless Networks,” Proc. missions to reach high throughput, but these 2013 Conf. Internet Measurement, 2013, pp. 455–468. mechanisms introduce additional latency. An 8. A. Ford et al., TCP Extensions for Multipath Operation important contribution could be given by multi- with Multiple Addresses, IETF RFC 6824, 2013; https:// connectivity through MP-TCP at the transport tools.ietf.org/html/rfc6824. layer, where path diversity offers promising 9. A. Ford et al., Architectural Guidelines for Multipath results, but a suitable CC algorithm has not TCP Development, IETF RFC 6182, 2011; https://tools been identified yet. .ietf.org/html/rfc6182.

18 www.computer.org/internet/ IEEE INTERNET COMPUTING TCP and MP-TCP in 5G mmWave Networks

10. C. Raiciu, M. Handley, and D. Wischik, Coupled Con- Michele Polese is a PhD student at the Department of Infor- gestion Control for Multipath Transport Protocols, mation Engineering of the University of Padova. His IETF RFC 6356, 2011; https://tools.ietf.org/html research interests are in the analysis and development /rfc6356. of protocols and architectures for 5G mmWave net- 11. R. Khalili et al., “MPTCP is Not Pareto-Optimal: works. Polese has an MSc in telecommunication engi- Performance Issues and a Possible Solution,” neering from the University of Padova. Contact him at IEEE/ACM Trans. Networking, vol. 21, no. 5, 2013, [email protected]. pp. 1651–1665. 12. Q. Peng, A. Walid, J. Hwang, and S.H. Low, “Mul- Rittwik Jana is a director of inventive science at AT&T tipath TCP: Analysis, Design, and Implementation,” Labs–Research. His research interests span intelligent IEEE/ACM Trans. Networking, vol. 24, no. 1, 2016, service composition of virtualized network functions pp. 596–609. (VNFs) using the Topology and Orchestration Speci- 13. B. Nguyen et al., “Towards Understanding TCP Perfor- fication for Cloud Applications (TOSCA) language, mance on LTE/EPC Mobile Networks,” Proc. 4th Work- model-driven control loop and automation in the Open shop on All Things Cellular: Operations, Applications, Network Automation Platform (ONAP), networked and Challenges, 2014, pp. 41–46. video streaming, and wireless communications and 14. R. Li, M. Shariat, and M. Nekovee, “Transport Protocols systems. Jana has a PhD in telecommunications engi- Behaviour Study in Evolving Mobile Networks,” Proc. neering from the Australian National University. Con- 2016 Int’l Symp. Wireless Communication Systems, tact him at [email protected]. 2016, pp. 456–460. 15. M. Polese et al., “Improved Handover Through Dual Michele Zorzi is with the Department of Information Connectivity in 5G mmWave Mobile Networks,” IEEE Engineering of the University of Padova. His present J. Selected Areas of Comm., special issue on Millime- research interests focus on various aspects of wireless ter Wave Communications for Future Mobile Networks, communications. Zorzi has a laurea degree and PhD in vol. 35, no. 9, 2017; https://arxiv.org/abs/1611.04748. electrical engineering from the University of Padova. 16. M. Polese, M. Mezzavilla, and M. Zorzi, “Performance He previously served as Editor-in-Chief of IEEE Wire- Comparison of Dual Connectivity and Hard Hando- less Communications and IEEE Transactions on Com- ver for LTE-5G Tight Integration,” Proc. 9th EAI munications, and is the current Editor-in-Chief of IEEE Int’l Conf. Simulation Tools and Techniques, 2016, Transactions on Cognitive Communications and Net- pp. 118–123. working. Contact him at [email protected]. 17. R. Ford et al., “A Framework for End-to-End Evaluation of 5G mmWave Cellular Networks in ns-3,” Proc. Work- shop on ns-3, 2016, pp. 85–92. 18. S. Choi and K.G. Shin, “A Class of Adaptive Hybrid ARQ Schemes for Wireless Links,” IEEE Trans. Vehicu- lar Technology, vol. 50, no. 3, 2001, pp. 777–790. 19. Korea Telecom (KT) Fifth-Generation Special Inter- est Group (5G-SIG), “5G.322, KT 5th Generation Radio Access, Radio Link Control protocol (5G pre-specifications),” stay connected. 2016. 20. M. Zhang et al., “Transport Layer Performance in 5G mmWave Cellular,” Proc. 2016 IEEE Conf. Computer Comm. Workshops, 2016, pp. 730–735. 21. J. Gettys and K. Nichols, “Bufferbloat: Dark Buffers in the Internet,” ACM Queue, vol. 9, no. 11, 2011, pp. 40: 40–40:54. | @ComputerSociety | @ComputingNow 22. M. Zhang et al., “The Bufferbloat Problem over Inter- mittent Multi-Gbps mmWave Links,” Nov. 2016; | facebook.com/IEEE ComputerSociety | facebook.com/ComputingNow https://arxiv.org/abs/1611.02117. 23. T. Azzino et al., “X-TCP: A Cross-Layer Approach for | IEEE Computer Society | TCP Uplink Flows in MmWave Networks,” Proc. 16th Computing Now Ann. Mediterranean Ad Hoc Networking Workshop, | .com/ieeecomputersociety 2017; https://arxiv.org/abs/1706.00904.

SEPTEMBER/OCTOBER 2017 19