You are on page 1of 6

Effect of Mode-Switching on TCP Short Flows

During D2D Communication in LTE-A Networks


Abhijan Bhattacharyya Bhaskar Sardar Debashis Saha
Embedded Systems and Robotics Dept. of Information Technology Management Information Systems
TCS Research and Innovation Jadavpur University Indian Institute of Management
Kolkata, India Kolkata, India Kolkata, India
abhijan.bhattacharyya@tcs.com bhaskargit@yahoo.co.in ds@iimcal.ac.in

Abstract—This paper investigates TCP performance in short- block of the smart cities. Proximity Service (ProSe)/device-to-
flows for device-to-device (D2D) communication in LTE-A device (D2D) communication, a proposed feature of LTE-A
networks. The target application is vehicle-to-vehicle (V2V) standardization under 3GPP, is promising to increase the
communication for traffic safety in intelligent transportation network capacity and is envisioned to extend the application of
systems (ITS). Through extensive simulation of real-life LTE to V2V communication in near future [9]. In this paper,
situations of ITS, our work highlights how even a very short-flow the aim is to transfer a few kilo bytes of data (which could have
over TCP is severely affected in an LTE-A network, when two been transmitted in one Round Trip Time or RTT had there
communicating end-points switch from D2D (or direct been no congestion control algorithm running) in the context of
communication mode) to infrastructure mode due to their
D2D communication between two mobile nodes, henceforth
mobility. Our analysis further proves that the existing solutions
to improve TCP performance for short flows over the
referred as User Equipment or UE, in two vehicles. While in
conventional Internet fail to cope up with such mode-switching. the middle of an ongoing TCP session, the nodes already
Finally, we propose a quick graceful solution, based on the communicating in D2D mode or direct mode (DM) may be
experimental results, to handle the unique problem, and discuss compelled, due to their mobility pattern, to switch back to
some more possible approaches for future research to circumvent infrastructure mode (IM) via conventional uplink/downlink
the related challenges of V2V communication. (UL/DL) through the LTE-A base station (i.e., eNB). This
switch over may be triggered by the eNB to maintain the best
Keywords—LTE-A, D2D, 5G, TCP, short-flow possible link quality. For instance, if two vehicular nodes,
communicating in DM, move away from each other, the Signal
I. INTRODUCTION to Interference plus Noise Ratio (SINR) may degrade over the
D2D link, called side link (SL), to such a level that the eNB is
Transmission Control Protocol (TCP) 1 is a connection forced to re-schedule the nodes to IM. This leads to a change in
oriented protocol for reliable delivery of data from source to connection state as the connection now has to be routed via the
destination. TCP’s congestion control mechanism ensures fair eNB, and this change-over needs certain additional processes
utilization of the available bandwidth. However, the design to be performed incurring additional delay and potential loss of
assumptions were: packet loss happens only due to congestion packets in flight. The loss in this case is neither due to
in the network and the usual exchange model is similar to FTP impairment in radio channel nor congestion in network; neither
or Telnet protocol which maintains long sessions of data is there a handover between eNBs. Hence, it is extremely
transfer. But, both the assumptions did not sustain with necessary to investigate TCP behavior under such conditions
advancement of the Internet. Firstly, Mobile Internet added so that corresponding corrective measures can be proposed.
several additional sources of loss ([1],[2]), including signal
impairments, and user mobility. Secondly, as WWW, powered The salient contributions of this paper include: (i)
with HTTP on TCP, became the most widely used application, simulation of a realistic situation for V2V communication to
dominant pattern for exchanges over the Internet turned out to show the effect of mode-switch for a very short TCP session;
be short transfers [3]. Subsequently, the original TCP the simulation setting presented in this paper can be reused as
congestion control mechanism, designed for preventing events part of a test-bed for future researches, (ii) experimental
like melting down [4], proved to be inefficient in terms of analyses based on the different metrics captured across the
resource utilization for short transfers [5]. The inefficiency protocol stack during investigation to reveal the cascading
became even more prominent with ever increasing bandwidth effect of the changing state of the access medium up to the
of communication links in both wired as well as wireless higher layer of the stack during mode-switch for a short
channels [6]. transfer, (iii) extensive experiments to show that standard
extensions of TCP, such as Selective Acknowledgement (SACK)
Though TCP short transfer has been well studied for [10] and Limited Transmit [11], are of no help in case of short-
conventional Internet, present work resurrects this old topic to flows under mode-switching; the causal analysis of the results
analyze it in the context of upcoming vehicle-to-vehicle (V2V) is presented to bring out the uniqueness of the problem that
communication for traffic safety in intelligent transportation TCP has to deal with in case of mode-switching, and (iv) our
systems (ITS) in smart cities. As observed in [7] and [8], smart recommendation for a novel application-aware cross-layer
cars combined with V2V technology is going to be the building solution for the potential application and network designers;
1
Transmission Control Protocol ., RFC 793

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE


plus several suggestions are put forward regarding the possible A. Problems with TCP Short Flow
future research directions. Slow-Start:TCP congestion control algorithms grow or
The next Section outlines a survey of existing literature. shrink the congestion window (cwnd) based on the feedback
Section III discusses background topics such as TCP short from the receiving end-point (through the acknowledgement or
transfer, brief idea about the LTE-A stack and resource ACK). They all start conservatively and take at least one RTT
allocation in D2D communication. Section IV presents the to grow the window size. However, in the attempt to avoid the
simulation experiments carried out and the results obtained risk of flooding the network, the network bandwidth remains
there from with requisite analyses. Section V discussed a largely underutilized. This becomes worse for a short flow. By
potential graceful solution. Section VI concludes the paper. the time a host ramps up to optimal transmission speed, the
entire data transfer could have been completed without
invoking any congestion control at all. Ever increasing
II. RELATED WORKS
communication bandwidth worsens the problem further [6].
The problem of TCP short flow over the Internet has been
identified long back [3]. Since then researchers have tried Loss Detection: TCP sender detects congestion based on
several way outs. For instance, in [12], the authors have two indications: timeout and three duplicate ACKs (dupacks).
proposed a new scheduling mechanism based on the prediction The short flows are not a good candidate for detecting loss
about the flow – short or long. Work in [5] has analyzed how using three dupacks. There may not be enough increase in
the network buffer size impacts the TCP short flows. In [13], cwnd to allow for three dupacks. Loss in short connection is
the authors have proposed TCP Fast Open to cater very short often detected only after a timeout. But it may happen that the
transactions over TCP. But all the above works belong to the timeout is detected after the entire data has been sent. As a
pre-LTE era, and hence do not talk about D2D communication. result, timeouts on short connections are not very effective
Recently, works in [14], [15] have dealt with TCP performance either in reducing the overall traffic or in stabilizing the
in LTE, and works in [16],[17] have addressed the aspects of network [12]. Occurrence of a loss in a short TCP transfer may,
mode-switching decisions in general. The work in [18] in effect, increase the transfer time manifold.
proposed new additions to the protocol stack to make mode-
switching faster for transferring delay-sensitive contents such B. Overview of LTE/ LTE-A Protocol Stack
as multimedia. However, to the best of our knowledge, there is LTE/LTE-A is an all IP system. The protocol architecture
no published work that investigates the impact of mode- can be broadly divided into: (i) Control Plane (responsible for
switching in LTE-A for TCP short-flows. Also, no work in the configuring the lower layers and handling radio-specific
literature till date has provided a graceful and quick solution to functionality depending on the state of the UE), and (ii)User
handle such TCP short-flows, as proposed in this paper. Plane (responsible for handling the application data) as shown
in Fig. 1. The LTE-A protocols actually form the Layer 2 of
The study that is closest to our work is the recent work in
the stack. However, in case of control plane, two components,
[19], which studied TCP performance for different congestion
Non Access Stratum (NAS) and Radio Resource Control
control algorithms in a simulated LTE-A environment under
(RRC), become part of Layer 3. In the user plane, IP packets
frequent mode-switching. However, they have considered a
traverse the Packet Data Convergence Protocol (PDCP) layer.
long TCP session for reliable transfer of a file of size 1 GB.
Here, they are ciphered and numbered to form PDCP Protocol
From the simulation set up in [19], it is evident that the nodes
Data Units (PDUs). The PDCP PDUs flow down to the Radio
will undergo mode-switching several times. This will definitely
Link Control (RLC) as RLC Service Data Units (SDUs). These
impact any long TCP connection with vehicles oscillating at
are kept in the RLC buffer. Each flow has one associated
very high speed. So, their purpose is completely different from
PDCP entity and one RLC entity. The PDCP layer assigns a
ours as we have already explained above. Furthermore, we also
Logical Connection Identifier (LCID), as per the signals
investigate how the different link level parameters change at
received during RRC connection set up, which uniquely
the point of mode-switching. This observation was not
identifies the logical channel instance within the scope of one
addressed in [19]. Moreover, we investigate the impact of a
pair of source and destination of a corresponding MAC flow.
standard solution like [11], which has been specifically
proposed for congestion control in case of TCP senders which There can be three RLC modes: transparent,
may have “only a limited amount of data to send.” Although unacknowledged, and acknowledged. Unacknowledged is
proposals for improving the LTE protocol performance using recommended in [9]. RLC performs segmentation/
cross-layer approach has been proposed in some literatures, concatenation of RLC SDUs on transmission, and reassembly,
none has addressed the same problem as ours. The cross layer duplicate detection and reordering of RLC PDUs on reception.
solution that we have recommended in this paper to address the Each time the RLC receives a MAC request for an RLC PDU,
identified problems is indeed unique as per our survey. it responds by de-queuing appropriate number of RLC SDUs
and then fragmenting or concatenating as necessary to form the
III. BACKGROUND
RLC PDU. The MAC layer adds the MAC header on top of the
As we all know, conventional TCP congestion control RLC PDU and forms a transport block (TB). The transmission
approach suffers from two key aspects in dealing with short of TB happens in sub-frames of 1ms duration and is paced at
transfers: slow start and loss detection. Transmission Time Intervals.
In the control plane, the Non Access Stratum manages the
session of the UE with the packet data network over IP
connectivity, and Radio Resource Control manages the responded by scheduling the UE in a future Transmission Time
signaling between UE and eNB and the configuration of Interval, and are re-iterated if unanswered post timeout. The
underlying layers. Beyond RRC, the information is processed scheduling is done based on the Buffer Status Report
by PDCP, RLC, MAC and Physical layer. (BSR).UEs periodically send the channel quality indicator
(CQI) to the eNB, which enables it to determine the
instantaneous Transport Block Size Index for the UEs. The
details of the process can be found in [9], [18], [19], [20].
However, an important question here is how to decide
whether a UE is to be allocated for D2D. This is yet another
open area for research. According to [9], the BSR from the
transmitting UE should indicate whether the UE is interested in
simple UL or a D2D resource allocation. But it is not specified
how the UE decides between UL and D2D, and how the
information is carried in BSR. We shall follow the approach
described in [20], which is the popular practice in place and
implemented in [21]. The BSR comes with a specific LCID
which indicates the flow direction. As per the implementation
in [21], a new flow ID for ‘D2D’ has been added to
differentiate between UL and D2D resource allocation requests.
A method for UE to decide the mode during channel access is
proposed in [22]. Let UE1 wants to transmit to UE2 in D2D
Fig. 1. The protocol architechture and data flow in LTE-A.
mode. According to [22], theUE1 first probes for UE2. The
probing process leads to a channel quality estimate for the
C. Overview of ProSe/D2D in LTE-A Network potential D2D link. Looking at the estimate, eNB can decide
As per the concept of proximity based service, if two UEs between UL and D2D, if UE1 registers its interest in D2D
are close to each other then they should be able to communication through the BSR.
communicate directly without routing via the eNB and the core How DM-IM switching affects communication: The
network. This way the effective network capacity can be mode-switching leads to certain change of connection-state in
increased. The different possible D2D scenarios are presented UEs as well as in eNB leading to losses in the network. As in
in [9]. Our work deals with D2D within a single cell, i.e., D2D [18], let us assume that two UEs ‘a’ and ‘b’ have been
enabled UEs under the coverage area of a single eNB. scheduled to communicate in DM (Fig. 2(a)). So, a PDCP
Use Case: Out of the several possible applications of D2D peering session, with its own sequence numbers and ciphering
in immediate and near future, our paper deals with possible parameters, will be established between the UEs. Two different
scenario from V2V communication. As envisioned in [8], the PDCP peering sessions have to be used (Fig.2(b)), when the
all IP based LTE-A, coupled with higher layer protocol communication is switched to IM. One peering is between ‘a’
offerings like TCP, is promising to be a technology of choice and the eNB (UL part), another between the eNB and ‘b’ (DL
for future V2V communication, which is going to be an part). Since these two PDCP sessions are independent,
important aspect of smart cities. connection state parameters (such as sequence numbers, and
ciphering/deciphering keys) will be unrelated and have to be
Mode-Switching in D2D: The radio resources to different different. Accordingly, the RLC PDU sequence numbers in SL
UEs in D2D mode are allocated by the corresponding eNB. are no longer valid in the UL/DL peering in IM. Consequently,
However, as pointed out in [20], the decision to do a mode- the data in the RLC buffer for the DM connection becomes
switch is one of the open research areas in unicast D2D invalid as the connection state has changed in IM. So, the
communication. While a transmission power based selection fragments in the old buffers at the respective UEs have to be
approach is proposed in [16], the approach in [17] performs discarded. This will incur a loss of data as explained in [18].
mode-switching based on channel quality. This approach has
been implemented in [21] too, going by the widely accepted
convention, and we also adapt it. The channel quality is
actually dependent on the positions of the nodes. So, the SL
channel quality may degrade at a point in time, when the two
UEs move away from each other. This is quite possible in a
V2V scenario. A path through the eNB may provide a better
channel under such circumstances, and the UEs should fall
back to connect via the eNB (i.e., IM) when the channel quality
in SL degrades. In such a scenario, the eNB is responsible for
re-assigning the radio resources.
Fig. 2. Mode-switching example.
D2D Resource Allocation: In an LTE-A system, UEs,
equipped with a Single Carrier-Frequency Division Multiple
Since TCP is a reliable connection oriented protocol, these
Access (SC-FDMA) receiver, request UL resources through a
lower-layer losses affect the TCP connection state. TCP tries to
Random Access Procedure (RAC). RAC requests are
recover from the loss, which in this case is due to neither timing diagram of Fig. 8. The mode-switch event is triggered
network congestion nor impairments in the wireless link. from eNB at 0.55 s. The communication stalls until UE1
However, due to its legacy design limitation, TCP has no other requests for grant at around 2.86 s. So, the overhead time due
choice but to consider the loss due to mode-switch as network to mode-switch is 2.31s.This observation is in sync with the
congestion and try to recover using the conventional cwnd in Fig. 6.Thus, even for very short transfer over TCP, a
congestion control mechanism. mode-switching can impact transfer and reduce the throughput
because TCP considers it as congestion (triggered by timeout)
IV. SIMULATION RESULTS AND ANALYSES and reacts with a slow-start phase.
We have used SimuLTE simulator ([20],[21]) on top of
OMNeT++2. Our primary goal of the simulation design is to
create a near-realistic situation within a single cell such that we
have a flow of events as presented in Fig 3. The biggest
challenge is to create a scenario, where, despite having to
transmit a very short data, a mode-switching can be triggered
realistically based on the CQI.
Initially, we place the UEs 100 meters away each from the
eNB (Fig. 4); they are 62 meters away from each other. Other
simulation parameters are given in Table I. Initially, we let the
UEs communicate in IM and DM. Figures 5(a) and 5(b) shows
the cwnd for IM and DM, respectively. The initial lag is due to
the resource allocation delay at eNB and the connection
establishment latency of TCP. As expected, the transfer
completes quickly without encountering any loss.

TABLE I. THE IMPORTANT PARAMETERS FOR SIMULATION.


Radio BW 10 MHz
LTE-A Number of cells 1
Parameters Number of UEs 2
D2D enabled Yes. For both UEs.
rwnd 50 KiB3
TCP MSS 536 bytes
TCP Version Reno
TCP SACK Disabled
Parameters TCP payloadsize (in KiB) 3/10/15/20
Min. RTO 1s
Max RTO 240s
RTO calculation As per RFC 6298

Then we let the UEs move away at a speed of 10 m/s (i.e.,


36 km/h) to trigger mode-switching. The cwnd graph in Fig. 6
reflects a congestion event. There is no impairment on the
radio channels. Congestion is also out of question in this case.
The loss in this case is due to mode-switch. This is confirmed Fig. 3. Pictorial description of the conceived flow of events.
by the physical layer captures at the NIC modules of the UEs
and the eNB (Fig. 7). The average UL CQI at the eNB is steady
due to its maintaining nearly constant distance from both the
UEs. In Fig. 7(b),the CQI at the transmitting node shows a dip
due to increasing distance between the UEs. The connection
disrupts close to the point where we experience a dip in the
cwnd. The right side of the graph shows the appearance of a
DL CQI establishing the fact that the UE has been rescheduled
in IM by the eNB. The appearance of the DL CQI coincides
with RTO event in Fig. 6(b). Fig. 7(c) shows the CQI variation
at the receiving node. The nature of the graph is evidently
similar to that of the transmitting node.
Fig. 4. Initial geo-location of the nodes in simulation.
We further analyze the timing of the events logged during
simulation. The important events are illustrated through the Next, we measure the impact on transmission time for
different short data transfers (Fig. 9), namely 10KiB, 15KiB
2
https://inet.omnetpp.org, https://omnetpp.org. and 20KiB. The increment in transmission delay due to mode-
3
KiB = Kibibyte; 1KiB = 1024 bytes.
switching grows with data size. The variance of the
transmission time also increases with increase in data size.
However, the rate of change in variance decreases with
increasing data size. Further investigations reveal that SACK
and Limited-Transmit optimizations fail to improve the
performance as shown in Fig. 10. SACK is not able to improve
because the transmit buffer has to be rejected in order to adjust
to the new connection state. The receive buffer also has to
reject anything that remains unacknowledged at the instant (a)
when mode-switching is triggered. Though Limited-Transmit
algorithm is based on optimization on dupacks, since the
buffers are out-dated at the receiving end as well as the
transmitting end, any delivery of future sequence also has to be
rejected.

(b)

(a)

(c)

Fig. 7. (a) The average UL CQI at the eNB. (b) Variation of CQI at the
transmitting node. (c) The CQI variation at the receiving node.

(b)

Fig. 5. The transmitter cwnd for transfer of 3KiB data (a) in IM, (b) in DM.

(a)

Fig. 8. Important events during simulation.

V. PROPOSED CROSS-LAYER SOLUTION


According to the results illustrated in Fig 9(a), the benefit
(b) in DM over IM is minuscule for very short flows. So, we
recommend that the transmitting UE should avoid mode-switch
Fig. 6. (a) The transmitter cwnd for transfer of 3KiB data in mode-switching. for such scenario. The UE should request resources in IM
(b) RTO event.
mode only. It should never express its interest in D2D
communication despite being capable to do the same. Thus, the the proposed solution in Section V. Also, it remains to be
flow direction in the LCID in the BSR remains ‘UL’. Going by investigated how the solution proposed in [18] fares in
the proposed recommendation, a negligible performance trade- comparison. It would be an interesting exercise to compare
off happens while the overall design remains frugal yet avoids between the proposed approach and [18] in future.
chances of degradation in goodput due to mode-switch. This
can be achieved by introducing an intelligent adaptive cross- REFERENCES
layer in between the application and the RRC (layer 3). If the
[1] Bello and S. Zeadally, “Communication Issues in the Internet of Things
application gets prior knowledge of the information length to (IoT)”, Next-Generation Wireless Technologies, Springer, 2013, pp.
be transmitted then the proposed layer may use this 189-219.
information before a scheduling grant. The adaptation layer [2] K. C. Leung, V. O. K. Li, “Transmission Control Protocol (TCP) in
would determine if the upcoming transfer is very short and wireless networks: issues, approaches, and challenges”, IEEE
then intimates the RRC to maintain the flow direction as “IM” Communications Surveys, Vol. 8., No. 4, 2006.
(as described in Section III) in the LCID. [3] H. Balakrishnan, V.N. Padmanabhan, S. Seshan, M. Stemm, R.H. Katz,
“TCP Behavior of a Busy Internet Server: Analysis and Improvements”,
in Proc. of IEEE Infocomm ’98, CA, USA, March 1998.
[4] S. Floyd, “RFC 2914 (Congestion Control Principles)”, September,
2000.
[5] Barakat C., Altman E. (2000) Performance of Short TCP Transfers. In:
Pujolle G., Perros H., Fdida S., Körner U., Stavrakakis I. (eds)
Networking 2000 Broadband Communications, High Performance
Networking, and Performance of Communication Networks.
NETWORKING 2000. Lecture Notes in Computer Science, vol 1815.
Springer, Berlin, Heidelberg
[6] N. Kartik, “TCP optimized for short flows”, Stanford University, June
2003, Link:
http://web.stanford.edu/class/ee384y/projects/download03/nitin3.pdf .
(a) [7] “Cars are the building blocks in a Smart City”, Link:
https://www.inmotionventures.com/cars-building-blocks-smart-city/,
July, 2016.
[8] A. Petrescu, “Intelligent Transportation Systems and the IETF”, IETF
Journal, July, 2016.
[9] 3GPP-TS 36.843 v12.0.1, “Study on LTE Device to Device Proximity
Services: Radio Aspects (Release 12)”, March 2014
[10] M. Mathis, RFC 2018, “TCP Selective Acknowledgement options”,
October, 1996.
[11] M. Allman, H. Balakrishnan, S. Floyd, RFC 3042, “Enhancing TCP's
loss recovery using limited transmit” , January, 2001.
(b) [12] K. Avrachenkov, U. Ayesta, P. Brown, E. Nyberg, “Differentiation
Fig. 9. (a) Effect of mode-switching on different sized short payloads. (b) between short and long TCP flows: predictability of the response time”,
Variance in exchange-time fordifferent payload sizes. INFOCOM 2004.
[13] Y. Cheng, et al, RFC 7413, “TCP Fast Open”, December 2014.
[14] D. Pacifico, et al., “Improving TCP performance during the Intra LTE
handover,” IEEE GLOBECOM, 2009.
[15] B. Nguyen, et al., “Towards understanding TCP performance on
LTE/EPC mobile networks,” Proc. of the 4th workshop on all things
cellular: operations, applications, & challenges. , 2014, pp. 41–46.
[16] J. Gu, S.J. Bae, B.G. Choi, M.Y. Chung, “Mode Selection Scheme
Considering Transmission Power for Improving Performance of Device-
to-Device Communication in Cellular Networks”, Proc. ICUIMC 2012,
20-22 Feb. 2012, Kuala Lumpur, Malaysia.
Fig. 10. Transmission time comparison among normal TCP-Reno, TCP-Reno [17] K. Doppler, Chia-Hao Yu, C.B. Ribeiro, P. Janis, "Mode Selection for
with SACK and TCP-Reno with Limited Transmit. SACK and Limited Device-To-Device Communication Underlaying an LTE-Advanced
Transmit do not show any benefit. Network," Proc. of WCNC 2010, pp.1-6, 18-21 Apr. 2010.
[18] G. Nardini, G. Stea, A. Virdis, D. Sabella, M. Careth, “Fast and agile
lossless mode-switching for D2D communications in LTE-Advanced
VI. CONCLUSION networks”, IEEE VTC, Spring, 2016.
In this paper, we have shown how mode-switching in D2D [19] G. Nardini, G. Stea, A. Virdis, “Performance evaluation of TCP-based
poses new challenge that remains to be solved for TCP short traffic over direct communications in LTE-Advanced”, IEEE VTC,
transfers in V2V communications. New research may bank Spring, 2016.
upon the findings in this paper and several directions of [20] A. Virdis, et al, “Modeling unicast device-to-device communications
with SimuLTE”, 1st International Workshop on Link- and System Level
ideation can appear, e.g., improving the scheduling algorithms, Simulations (IWSLS), July, 2016.
modification of the protocol mechanism during mode-switch, [21] SimuLTE – LTE User Plane Simulator for OMNeT++ and INET
special optimization on TCP to deal with mode-switch, use of (http://simulte.com).
predictive approach to schedule mode-switching intelligently, [22] US 2012/0184306 A1, “RANDOM ACCESS PROCESS REUSING
etc. Further implementation and testing is to be carried out for FOR D2D PROBING IN CELLULAR-AIDED D2D NETWORKS”.