Академический Документы
Профессиональный Документы
Культура Документы
Rethink Hybrid
Automatic Repeat reQuest Design for 5G:
Five Configurable Enhancements
Klaus I. Pedersen, Saeed R. Khosravirad, Gilberto Berardinelli, and Frank Frederiksen
ple for a downlink HARQ process, where the Finally, it should be noted that the results in practice also vary depend-
ing on the payload size,
timing elements such as the processing times, in Table 1 are for the case with zero fronthaul i.e. allowing even shorter
the TTI duration, the ACK/NACK duration and latency. For cases with non-zero fronthaul processing times for small
the network architecture delay influence on the latency (as illustrated in Fig. 1), the number of transport block sizes.
required SAW processes increases even further. as well as the maximum number of transmissions
Similarly, the use of multi-stream MIMO also Kmax naturally affect the performance. A con-
calls for a proportional increase in the number of figuration with BLER1 = 1 percent and Kmax =
SAW processes. 2 leads to the selection of a low-order modula-
tion and coding scheme and low average latency,
Early Feedback for LLC while a configuration with BLER1 = 30 percent
The processing time at the UE could become a and Kmax = 8 leads to more aggressive rates at
significant bottleneck for achieving the challeng- the expense of higher average latency. Using an
ing LLC targets set for 5G. In current LTE receiv- early feedback facilitates using aggressive rate
ers, it is estimated that around 60 percent of the configurations without compromising the latency.
processing time is spent for turbo decoding, while For instance, it is estimated in [7] that the same
the remaining time is spent for OFDM processing, (or lower) average latency can be achieved with
equalization, demodulation, estimation of soft val- BLER1 = 10 percent and Kmax = 4 with early feed-
ues for channel decoding, and so on. Predicting back, and configuration with BLER1 = 1 percent
whether the decoding will be successful prior to and Kmax = 2 with regular feedback. Of the two
the decoding itself, would enable the UE to antic- configurations, the former benefits from higher
ipate its feedback transmission while running the spectral efficiency.
full decoding operation in parallel. This will result Figure 2 shows the probability of having
in reduced HARQ RTT. The prediction can be a larger HARQ latency than the value on the
performed upon reception of the TTI, or even x-axis, assuming different feedback types. The
a portion of the TTI, for example, only a limited feedback transmission time is set to 0.0625 ms
number of OFDM symbols. (i.e., corresponding to a UE in good coverage
The decoder outcome can be predicted from conditions). A downlink TTI duration of 0.125
the log-likelihood ratios (LLR) of its input bits, ms is considered, as well as a (BLER1 = 10 per-
which are obtained from the estimation of soft cent, Kmax = 4) configuration, with 0.7 combin-
values during the demodulation of the equal- ing efficiency. The processing time at the base
ized data symbols. The proposal in [7] is based station and terminal is set to 0.125 ms. A false
on estimating the uncoded bit error rate (BER) negative (FN) rate of 2 percent and 10 percent
from such likelihood ratios, and mapping such is assumed. A FN rate of 10 percent is a rath-
estimate to reference coded BER curves of a spe- er pessimistic case, while the 2 percent case is
cific code. Though the performance is evaluated compliant to the initial findings in [7]. The pres-
for turbo codes, the principle can be generalized ence of FNs increases the probability of a first
to other coding solutions. Simulation results show retransmission, but achieves improvements of
the probability of obtaining a correct prediction the tail of latency distribution. For instance, at
in more than 90 percent of the considered cases. 10 –5 probability, latency of the early feedback
However, the average rate of erroneous estimates with 2 percent FN rate is still lower than the one
is not negligible. False positives occur when an for regular feedback. Improving the reliability of
early ACK is generated for a codeblock that is not the decoder prediction by further minimizing the
going to be correctly decoded, while false neg- probability of such wrong estimates (including
atives occur when an early NACK is generated false positives) is a valid research direction for an
for a successfully decoded codeblock. False pos- enhanced HARQ concept.
itives could result in higher layer retransmissions.
This leads to a latency increase. Conversely, false Adaptive-Length Redundancy Matching with
negatives may cause unnecessary retransmissions, Enriched Feedback
which affect the throughput but have a limited The feedback message for the HARQ process
impact on the latency. False positives are consid- is traditionally a simple Boolean type value that
ered more critical for LLC cases. The solution in conveys ACK or NACK, in order to minimize the
[7] shows that their occurrence can be minimized feedback overhead. This is inefficient compared
at the expense of an eventual higher rate of false to the case where more accurate decoder state
negatives. information (DSI) is available at the transmitter
It is worth mentioning that the use of an early node. The DSI conveys information on “how
feedback can also be beneficial for spectral effi- close” the receiver was at being able to correctly
ciency improvement. The HARQ setting for the decode a failed HARQ transmission. In the pres-
block error rate target at first transmission (BLER1) ence of DSI at the transmitter, an optimal decision
0.7
line with the 3GPP simulation assumptions [4]. A
0.6
20 MHz carrier bandwidth at 2 GHz with an LTE-
like physical layer numerology is assumed. Fig. 3 0.5 Fixed size
shows the cumulative distribution of the relative 0.4 Ideal feedback
size for the first retransmission compared to the 3-bit floor
initial transmission size. The quantization options 0.3 3-bit ceil
2-bit floor
labelled “floor” and “ceil” choose a value that 0.2 2-bit ceil
is the closest smaller value, or the closest larger
0.1
value, to the calculated NACMI at the decoder,
respectively. The results are generated for BLER1 0
10 -2 10 -1 10 0
= 10 percent and BLER2 = 0.1 percent. For the First retransmission size ratio
case of ideal feedback (i.e., no quantization of
the NACMI), the savings in the required radio FIGURE 3. The ratio of the first retransmission relative to the initial transmission
resources for the HARQ retransmissions are up size.
to 90 percent in ~70 percent of the failed initial
transmission samples when compared to fixed
size HARQ operation. It is shown that quantizing be up to 100 MHz [5], while it may reach val-
the feedback message to 2–3 bits for reporting ues of GHz for millimeter wave bands. Hence,
the NACMI can potentially save up to 90 percent scheduling a user over the full transmission
of the resources used for the first retransmission bandwidth results in a very large transport block
for 35–75 percent of the cases, depending on the (TB) size (TBS). As in LTE, a TB, if larger than a
quantization method. The reduced retransmission certain number of information bits (e.g., 6144
resource size in this case enhances the overall bits), gets segmented into several code blocks
downlink cell throughput by ~8 percent at the (CBs) so that each CB size (CBS) is smaller or
expense of the higher multi-bit HARQ feedback equal to that limit [15, Sec. 5]. On the receiv-
overhead in the uplink. Further improvements of er side, each CB will be decoded separately,
the link adaptation and coding methods as well and a traditional single-bit ACK is sent only if all
as the joint coding/decoding of multi-message the CBs in the TB are decoded correctly; oth-
HARQ solutions [8] to accommodate adap- erwise, the feedback is a NACK. As a result, a
tive-length redundancy matching HARQ are other single failed CB will mean that the TB is marked
research topics that should be considered for the as erroneous even though it would suffice to
5G era. retransmit the failed CB only. For a failed large
TB transmission and a subsequent NACK, the
Enhancements for following two extreme scenarios can be con-
sidered:
Large Transport Block Size • One or a few CBs fail in the decoding process.
The larger transmission bandwidth for 5G The most attractive information for the trans-
naturally provides larger data rates, but also mitter in this scenario is to know which CBs
challenges for the HARQ design. The carrier have failed and only perform a retransmission
bandwidth for 5G below 6 GHz is expected to for the failed CBs.
Low latency • Early HARQ feedback to reduce the HARQ round trip • Early decoder prediction even before having received the complete transport block.
communication (LLC) time (RTT) • Prediction based on log-likelihood ratios (LLR).
• Multi-bit feedback information on “how close” the receiver was at being able to correctly
decode a failed HARQ transmission.
• Optimize resource usage for HARQ retransmissions.
Variable block length • Expressing the amount of extra information that is needed for the successive HARQ
• Benefit from richer HARQ feedback as compared to
HARQ retransmission.
traditional Boolean feedback.
• Multi-bit feedback based on quantized version of the normalized accumulated mutual
information (NACMI).
• Support for low cost and low energy MTC devices with
• Only few downlink HARQ SAW processes per device.
minimum memory.
• Synchronous HARQ to lower the energy consumption.
Low cost MTC • Use cases with non-scheduled uplink access, NOMA,
• Further research on HARQ schemes for non-scheduled access with NOMA and/or
SCMA is of relevance.
SCMA is recommended.
• Half-duplex devices.
TABLE 2. Summary of proposed HARQ enhancements for the 5G-era.
ther, the concept of feedback is expected to take a [4] 3GPP, “Technical Report (TR) 38.913. Study on scenarios
and requirements for next generation access technologies,”
leap forward from its Boolean nature and provide Mar. 2016.
further information for optimizing the retransmis- [5] K. I. Pedersen et al., “A Flexible 5G Frame Structure Design
sions. Transmission in very large bandwidth calls for Frequency-Division Duplex Cases,” IEEE Commun. Mag.,
for the use of efficiently designed and acknowl- vol. 53, no. 3, Mar. 2016, pp. 53–59.
[6] S. Khosravirad et al., “Enhanced HARQ Design for 5G Wide
edged multiple code blocks per transmission time Area Technology,” IEEE Proc. Vehicular Technology Conf.
interval. Finally, it is advisable to have small-size (VTC), May 2016.
synchronous HARQ for low energy mMTC devices [7] G. Berardinelli et al., “Enabling Early HARQ Feedback in
with few SAW processes. It is proposed that the 5G Networks,” IEEE Proc. VehicularTechnology Conf. (VTC),
May 2016.
network is in charge of enabling (i.e., configuring) [8] A. Benyouss et al., “Joint Coding/Decoding for Multi-Mes-
the supported HARQ enhancements whenever sage HARQ,” IEEE Proc. Wireless Commun. Networking
needed, for example, depending on the carried Conf., June 2016.
service. This is essentially allowing a user-centric [9] F. Boccardi et al., “Why to Decouple the Uplink and Down-
link in Cellular Networks and How to Do It,” IEEE Commun.
HARQ configuration in coherence with its service Mag., vol. 53, no. 3, Mar. 2016, pp. 110–17.
requirements as a means to optimize the end-to- [10] H. Ding et al., “On the Performance of HARQ-IR over
end performance. Nakagami-m Fading Channels in Mobile Ad Hoc Networks,”
In addition to the HARQ enhancements iden- IEEE Trans. Veh. Technol., Aug. 2016.
[11] Y. Long et al., “A Novel HARQ Scheme for SCMA Sys-
tified in this study, there are further opportuni- tems,” IEEE Wireless Commun. Lett., vol. 5, no. 5, pp. 452–
ties for improvements to be explored in future 55, Oct. 2016.
research. These include, but are not limited to, [12] S. R. Khosravirad et al., “Rate Adaptation for Cooperative
enhancements that are optimized for new forward HARQ,” IEEE Trans. Commun., vol. 62, no. 5, May 2014,
pp. 1469–79.
error correction schemes such as polar codes and [13] S. R. Khosravirad et al., “HARQ Enriched Feedback Design
low density parity codes (LDPC), HARQ schemes for 5G Technology,” IEEE Proc. Vehicular Technology Conf.
tailored for NOMA and SCMA use cases, as well (VTC), Sept. 2016.
as custom designed HARQ schemes for connec- [14] K. I. Pedersen et al., “A Flexible 5G Wide Area Solution
forTDD with Asymmetric Link Operation,” IEEE Wireless
tivity use cases such as D2D and multi-hop wire- Commun., 2016.
less backhaul links. [15] 3GPP, “Technical Specification (TS) 36.212. LTE; Evolved
Universal Terrestrial Radio Access (E-UTRA); Multiplexing
References and Channel Coding (Release 13),” April 2016.
[1] D. Chase, “Code Combining: A Maximum-Likelihood Decod-
ing Approach for Combining an Arbitrary Number of Noisy
Packets,” IEEE Trans. Commun., vol. COM-33, no. 5, May Biographies
1985, pp. 385–93. Klaus I. Pedersen (klaus.pedersen@nokia-bell-labs.com) received
[2] L. Szczecinski et al., “Rate Allocation and Adaptation for the M.Sc. degree in electrical engineering and the Ph.D. degree
Incremental Redundancy Truncated HARQ,” IEEE Trans. from Aalborg University, Aalborg, Denmark, in 1996 and 2000,
Commun., vol. 61, no. 6, June 2013, pp. 2580–90. respectively. He is currently leading the Nokia Bell Labs research
[3] A. Larmo et al., “The LTE Link-Layer Design,” IEEE Commun. team in Aalborg, and he is a part-time professor at Aalborg
Mag., vol. 47, no. 4, pp. 52–59, April 2009. University in the Wireless Communications Network (WCN)