MIMOPack

A HPC Library for MIMO Communication Systems

Carla Ramiro Sánchez

, Antonio M. Vidal Maciá

and Alberto Gonzalez Salvador

Dpto. De Sistemas Informáticos y Computación (DSIC), Universitat Politècnica de València, Valencia, Spain

Instituto de Telecomunicaciones y Aplicaciones Multimedia (iTEAM), Universitat Politècnica de València, Valencia, Spain

1 RESEARCH PROBLEM

Nowadays, several communication standards are

emerging and evolving searching higher

transmission rates, reliability and coverage. This

expansion is primarily driven by the continued

increase in consumption of mobile multimedia

services due to the emergence of new handheld

devices such as smart phones and tablets.

One of the most significant techniques employed

to meet these demands is the use of multiple transmit

and receive antennas, known as Multiple-Input

Multiple-Output (MIMO) systems (Paulraj et al.,

2004). The use of this technology allows to increase

the transmission rate and the quality of the

transmission through the use of multiple antennas at

the transmitter and receiver side.

MIMO technologies have become an essential

key in several wireless and broadband standards

such as Wireless Local Area Network (WLAN),

Wordlwide interoperability for Microwave Acces

(WiMAX), Long Term Evolution (LTE) and Next

Generation Handheld (DVB-NGH), for the reception

of digital terrestrial television (TDT) for handheld

devices. These technologies will be incorporated

also in future standards, therefore is expected in the

coming years a great deal of research in this field.

Clearly, the study of MIMO systems is critical in

the current investigation, however the problems that

arise from this technology are very complex. The

detector is present in the receiver side and is the

responsible for recover the transmmitted signals

(which are affected by the channel fluctuations) with

the maximum reliability. This step becomes in many

cases the most complex stage in the communication.

Another important factor that affects the

performance of a MIMO system is the number of

transmit and receive antennas, because as the system

grows the communication data processing becomes

more complicated. Although the number of antennas

currently allowed in the standards is not large, it is

expected that in the near future more than 100

transmit antennas will be used (Rusek, 2013),

(Lamare, 2013). All the above reasons motivate the

search for high-throughput versatile receiver

implementations capable to be reconfigured and

scalable with the system parameters.

The High Performance Computing (HPC)

systems, and specifically, modern hardware

architectures as multicore and manycores (e.g

Graphic Processing Units (GPU)) are playing a key

role in the development of efficient and low-

complexity algorithms for MIMO transmissions.

Proof of this is that the number of scientific

contributions and research projects related to its use

has been increased in the last years (Wu, 2011),

(Nylanden, 2010), (Falcao, 2009). Also, some high

performance libraries have been implemented as

tools for researchers or companies involved in the

development of future communication standards, for

example: IT++ library based on the use of some

optimized libraries for multicore processors (IT++,

2013) or the Communications System Toolbox of

Matlab (MathWorks, 2014) which use GPU

computing. However, there is not a library able to

run on a heterogeneous platform using all the

available resources whenever possible.

In view of the high computational requirements

in MIMO research and the shortage of tools able to

satisfy them, we have made a special effort to

develop a library to ease the development of

adaptable parallel applications in accordance with

the different architectures of the executing platform.

The library, called MIMOPack, aims to implement

efficiently, using parallel computing, a set of

functions to perform some of the critical stages in

MIMO communication systems. This library can be

run over the last generation of machine architectures

(e.g GPUs and multicore), or even simultaneously,

since it is designed to use on heterogeneous

machines to exploit the whole computational

capacity thus reducing the response time of the most

complex problems.

MIMOPack, may allow industrial and academic

researchers to include more complex algorithms in

Sánchez C., Maciá A. and Salvador A..

MIMOPack - A HPC Library for MIMO Communication Systems.

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

their simulations and obtain its results faster.

Moreover, it can be run in a wide range of

architectures even many of them callable from

MATLAB increasing the portability of developed

codes between different computing environments.

2 OUTLINE OF OBJECTIVES

The use of MIMO technology has had enormous

repercussions in today's telecommunications systems

and certainly it will do in the near future. The

benefits offered are achieved at the expense of an

increase in the material costs to deploy multiple

antennas at both the transmitter and the receiver, and

also at the expense of additional complexity at the

receiver end of the MIMO system. For that reason,

signal detection has been the subject of intense study

during the last decade and the search for high

throughput practical implementations also scalable

with the system size remains essential today.

The grand challenge is to develop fast algorithms

to optimize the design and the validation process of

new MIMO schemes and technologies. Then, this

work aims to contribute to meeting this goal. This

section describes some particular motivations

leading towards the high performance library design.

The use of HPC systems brings big benefits, but

it will also poses big challenges. These systems

introduce asymmetries and heterogeneities that

complicate the development of efficient algorithms.

In the recent years, a large variety of machine

architectures have appeared, in view of this situation

researchers of the scientific community are obliged

to write the codes in different programming

languages and consider many details of the

architecture to use efficiently the whole target

system. Therefore, this high performance computing

library is essential to facilitate the implementation of

scientific codes on a widespread range of

architectures.

Furthermore, there are several important

companies implicated in the development of new

communication standards. In the standardization

process different entities are involved such as:

administrations, network operators, manufacturers,

users, research bodies, universities, consultancy

companies, partnerships and others (ETSI, 2014).

The objective of a standard is to provide a set of

rules, guidelines or characteristics ensuring the

interoperability between systems developed by

different manufacturers.

Figure 1: Nature and scope of the thesis.

These manufacturers work to propose their own

technologies with the aim that it will be approved as

standard. This will allow to enforce their intellectual

property rights over its competitors, who will must

obtain the corresponding license to incorporate the

technology adopted as standard into their products.

In many cases, simulation is the only way to get

these proposals but these simulations often involve a

high computational burden, since they try to

simulate the transmission of large amount of bits in

order to obtain results close to than would be

obtained in a real transmission. Normally these

simulations require weeks or even months to be

completed. Thereby, MIMOPack may allow the

launch of large simulations, opening the door for

industrial researchers to analyze its technologies

faster than its conventional simulation and obtain

more patent opportunities than their potential

competitors. Taking into account the above

presented motivations, the main scope of this thesis

is the following (see Figure 1):

 To develop an efficient library of functions able

to perform some of the most important and

complex stages in a MIMO communication as

preprocessing, precoding, detection and

decoding.

 To contribute with high-throughput

implementations of functions using parallel

processing and to evaluate them in terms of

speedup, bit error rate and throughput and

compare them with other existing

implementations.

 To facilitate to the programmer the

implementation of codes on a wide range of

architectures, incrementing the portability of

codes between different computing platforms by

using common interfaces for all the considered

environments. This approach simplifies the use

of the library, regardless of the machine where it

will be executed.

PECCS2015-DoctoralConsortium

 To develop a set of functions highly configurable

able to be executed with different simulation

parameters and different computational

resources.

3 STATE OF THE ART

Programming applications over HPC systems allows

to reduce the execution time of complex problems,

but it use leads serious programming difficulties.

The programming challenge involves the developers

to know different programming languages and the

characteristics of the architecture in depth. In this

sense the high performance libraries become in

valuable tools for specialists of a particular field,

since it facilitate the development of scientific

codes.

Some software companies have already released

to the market various libraries, these libraries not

only facilitate the preparation of code but also

exploit the huge computing capabilities of new

architectures to accelerate and optimize these

implementations. There are several libraries with

extensive backgrounds and acceptance in the

scientific community, most of them in numerical

linear algebra. For example the Linear Algebra

PACKage (LAPACK) designed to run efficiently on

shared-memory vector and parallel processors

(Anderson, 1999). There are also some versions of

them, which provide implementations of basic

mathematical functions on GPUs as for example

CULA (CULA, 2014) or MAGMA (Tomov, 2011).

In the field of communication systems

applications, few tools or HPC libraries are

available. Nevertheless, we can find two remarkable

libraries:

 Communications System Toolbox provides

algorithms for designing, simulating, and

analyzing communications systems. These

capabilities are provided as MATLAB functions,

MATLAB System objects, and Simulink blocks.

The system toolbox enables source coding,

channel coding, interleaving, modulation,

equalization, synchronization, and channel

modeling. Also allows analyze bit error rates,

generate eye and constellation diagrams, and

visualize channel characteristics. Although this

software is excellent and widely used by the

scientific community, nowadays just a small set

of functions are prepared to use parallel

computing with GPUs.

 IT++ Library: is a C++ library of mathematical,

signal processing and communication classes and

functions. The kernel of the library consists of

generic vector and matrix classes, and a set of

accompanying routines. IT++ makes an

extensive use of existing open-source or

commercial libraries for increased functionality,

speed and accuracy (e.g BLAS, LAPACK,

FFTW, ATLAS, ACML, etc). However, this

library is oriented to its exclusive use on

multicore machines; it does not have support to

use GPUs.

4 METHODOLOGY

The methodology to be followed in this research

work is summarized in the following tree phases.

4.1 Conceptual Phase

This phase includes from conception of the research

problem to the concreteness of objectives.

 Find information about the current MIMO

systems and techniques used in different modules

existent in the communication process (Channel

coding and decoding, preprocessing, precoding,

hard/soft detection).

 Bibliographic review on efficient algorithms and

high performance libraries for MIMO

communication systems.

 Enumerate the main objectives and detect the

main motivations leading towards the library

implementation.

4.2 Methodological Phase

This phase includes the algorithms design and its

implementation on high performance hardware.

 Study of the performance of the algorithms

implemented sequentially to establish which

parts of the code are the computationally most

expensive.

 Design of algorithms: The parallelization

opportunities of each algorithm will be

discussed, and also the possibility of its

implementation with heterogeneous multicore +

heterogeneous multi-GPU programming.

 Implement the algorithms designed in the

previous stage using OpenMP (for multicore

systems) and CUDA for GPU devices.

Moreover, will be explored the possibility of

using numerical linear algebra libraries and

MIMOPack-AHPCLibraryforMIMOCommunicationSystems

optimized as LAPACK, BLAS, etc in some parts

of the code.

4.3 Empirical Phase

This phase involves validation, analysis and

interpretation of the results obtained with our

implementations.

 Analysis of different parameters to obtain the

best performance in the design of each module.

For example: number of threads, type of

distribution of tasks (static or dynamic), and

thread block size.

 Validation stage: Unit testing for each function

implemented above. These tests allow us to

ensure that our code work properly.

 Analyze the performance based on speedup, Bit

Error Rate and throughput.

 Perform the same computational analysis on

several systems, checking if the library is

portable and has a good performance on any type

of hardware platform. For example, different

types of machines GPUs, other numbers of

processors, etc.

 Optimization performance and efficiency of

various library functions. The optimization is

carried out either if the speedup was not

expected, or when necessary to optimize the code

for certain fixed parameters specified in a

particular standard.

5 EXPECTED OUTCOME

As a result of the research work described in

previous sections, a C library will be able for

simulation of MIMO communication systems. The

library will include tools and functions to perform

some of the critical stages in the communication

process, which can be executed over different types

of architectures to ease the development of adaptable

parallel applications in accordance with the different

architectures of the executing platform:

 Sequential processor

 Multi-core processor

 GPU and Multi-GPU

 Heterogeneous (multicore and GPUs)

For each simulation the library shall allow to

configure the following MIMO systems parameters:

 Number of transmitter/receiver antennas

 Signal to Noise Ratio (dB)

 Modulation

 Number of signal vectors to be transmitted

 Variation of the channel

Each function will be executed with the

configuration selected by the user:

 Number of CPU threads

 Number of GPUs

 Quantity of workload for GPUs and CPU

 Use or not MKL library whenever possible

6 STAGE OF THE RESEARCH

The library is composed by several modules. In

Figure 1 we can see the basic simulation chain

through the library headers, as we can see the library

allows the user to assign the type and the number of

resources to use during the execution and get the

simulation results to calculate some statistics (e.g

simulation time, Bit Error Rate, Symbol Error Rate

or throughput).

Figure 2: Simulation chain through the MIMOPack library

modules.

6.1 Software Package Functions

The library is continuously growing; the updated

release collects a set of functions to compute the

most important and complex stages in a MIMO

communication which are described in this section.

6.1.1 Hard-output Detectors

If nearly optimal data detection is desired, the

detector becomes often the most computationally

expensive algorithm within a MIMO receiver. The

detector is responsible of processing the received

mixture of signals affected by the channel in order to

recover the transmitted data with the accuracy

required by the considered application. This issue

motivates the search for high-throughput MIMO

hard-output detectors capable to be reconfigured and

scalable with the system parameters. The hard-

PECCS2015-DoctoralConsortium

output detectors currently available in the library are

listed below:

 MLE: MLE-Exhaustive (Agrell, 2002).

 SESD: Schnorr Euchner Sphere Decoder

(Schnorr, 1994).

 ZF-SIC: Zero Forcing SIC (Berenguer, 2003).

 HFSD: Hard Fixed Sphere Decoder (Barbero,

2008).

 K-BEST: K-Best Decoder (Guo, 2006).

 ASD: Automatic Sphere Decoder (Su, 2005).

Also, some strategies such as channel matrix

preprocessing techniques can be used in order to

decrease the computational cost of data detection

and are also implemented in MIMOPack.

6.1.2 Soft-Output Detectors

Error control coding ensures the desired quality of

service for a given data rate and is necessary to

improve reliability of MIMO systems. Therefore, a

good combination of detection MIMO schemes and

coding schemes has drawn attention in recent years.

The most promising coding schemes are Bit-

Interleaved Coded Modulation (BICM) (Caire,

1998). At the transmitter the information bits are

encoding using an error-correction code. The soft

demodulator provides the reliability information in

form of real valued log-likehood ratios (LLR). These

values are used by the channel decoder to make final

decisions on the transmitted coded bits.

Nevertheless, these sophisticated techniques produce

a significant increase in the computational cost and

require large computational power. The following

Soft-Output detectors have been implemented:

 Soft Fixed Sphere Decoder (Barbero, 2008).

 Fully Parallel Fixed-Complexity Detector

 Max-Log Detector (Müller, 2002).

 ML Optimum Detector (Müller, 2002).

6.1.3 Multiuser MIMO Precoding

In multi-user (MU) MIMO transmissions,

specifically, in the downlink scenario, a base station

equipped with multiple antennas transmits

information to several independent users. The

detection process becomes more complicated due to

the absence of cooperation between them. In order to

simplify the detector complexity at the receiver side,

several precoding techniques were devised by

various authors and have been included in the library

(Windpassinger, 2004), (Yao, 2002):

 Zero-Forcing Precoding

 Tomlinson-Harashima.

 Lattice Reduction Aided Tomlison-Harashima

 Enhanced Lattice Reduction Aided

6.2 Preliminary Performance Results

In order to asses the performance of the library, we

have evaluated the execution time of a set of Hard-

Output detectors. The tests are executed on a

platform with one Nvidia Tesla K20Xm GPU with

14 SM, each SM including 192 cores. The core

frequency is 0.73 GHz. The GPU has 5GB of

GDDR5 global memory and 48KB of shared

memory per block. The installed CUDA toolkit is

5.5. The Nvidia card is mounted on a PC with two

Intel Xeon CPU E5-2697 at 2.70 GHz with 12 cores

and hyperthreading activated.

We consider as simple simulation example of a

MIMO system with 6 transmit and receive antennas,

16-QAM symbol alphabet and a constant channel

during the entire simulation. The speedup is defined

as the ratio between the computational time resulting

of executing the simulation of 50000 signals on the

sequential CPU (with one OpenMP thread) and the

time to execute the same simulation on a multicore

and GPU system.

By linking the code with MIMOPack library, we

can easily reduce the execution time of our

simulations significantly. This execution time can be

decreased even more using the heterogeneous mode

(i.e multicore and GPU simultaneously).

Table 1: Runtime of Hard-Output MIMOPack detectors

with different library configurations.

Detector

Runtime (sec)

Sequential 48 Threads GPU

MLE 1.48 · 10

7.97 · 10

5.69 · 10

SESD 89.08 3.88 12.80

L = 1

HFSD 0.45 0.13 0.34

2-BEST 0.48 0.07 0.50

4-BEST 0.82 0.14 0.74

16-BEST 3.31 0.26 1.64

L = 3

HFSD 64.70 3.24 3.71

2-BEST 15.98 0.81 3.86

4-BEST 16.49 0.84 4.15

16-BEST 20.45 0.89 5.02

64-BEST 37.07 1.51 7.01

As we can see in Table 1, we have a good

speedup for the parallel versions for all kind of

detectors. MLE GPU version exhibits a good

MIMOPack-AHPCLibraryforMIMOCommunicationSystems

performance even better than that obtained with

multicore version. However, due to the low

complexity and the non parallel pattern of other

detectors especially of the suboptimal ones (ZF-SIC

and HFSD) methods, the OpenMP version obtains

better performance than CUDA version.

This gain gradually disappears when the

complexity of the detector increases, for example

with the number of levels to be fully expanded (L)

or increasing the number of survivors (K) to be

computed in each level in the K-BEST detector.

The variety of detectors with mixed complexities

and performances allows to cover multiple use cases

with different channel conditions and scenarios such

as massive MIMO. Moreover, parallel

implementations allow the execution of large

simulations over different architectures thus

exploiting the capacity of the modern machines.

7 CONCLUSIONS

This thesis is focused in the development of a high

performance library for MIMO communications

systems which aims to

provide a set of routines

needed to perform the most complex stages in the

current wireless communications. The proposed

library has three important features: portable,

efficiently and user friendly. Results obtained with

the efficient hard-output detectors presented in this

paper demonstrate that MIMOPack library may

become in a very useful tool for companies involved

in the development of new wireless and broadband

standards which need obtain results and statistics of

its proposals quickly and also for other researchers

making easier the implementation of scientific

codes.

ACKNOWLEDGEMENTS

This work has been supported by SP20120646

project of Universitat Politècnica de València, by

ISIC/2012/006 and PROMETEO FASE II 2014/003

projects of Generalitat Valenciana; and has been

supported by European Union ERDF and Spanish

Government through TEC2012-38142-C04-01.

REFERENCES

Paulraj A.J., Gore D.A., Nabar R.U., Bölcskei H., 2004.

An overview of MIMO communications-a key to

gigabit wireless. Proceedings of the IEEE, 92(2):198-

218.

Rusek F., Persson D., Lau B., Larsson E., Marzetta T.,

Edfors O., Tufvesso F., 2013. Scaling Up MIMO:

Opportunities and Challenges with Very Large Arrays.

IEEE Signal Processing Magazine, 30(1):40-60.

Lamare R.C., 2013. Massive MIMO Systems: Signal

Processing Challenges and Research Trends. URSI

Radio Science Bulletin.

Wu M., Sun Y., Gupta S., Cavallaro J., 2011.

Implementation of a high throughput soft MIMO

detector on GPU. Journal of Signal Processing

Systems, 64(2):123-136.

Nylanden T., Janhunen J., Silven O., Juntti M., 2010. A

GPU implementation for two MIMO-OFDM detectors.

International Conference on Embedded Computer

Systems.

Falcao G., Silva V., Sousa L., 2009. How GPUs can

outperform ASICs for Fast LDPC decoding.

International Conference of Supercomputing.

IT++., 2014. IT++ User’s guide. http://itpp.sourceforge

.net /4.3.1/users_guide.html.

MathWorks., 2014. Communication System Toolbox.

User’s guide Version 6.5. http://jp.mathworks.com/

help/pdf_doc/comm/comm.pdf.

ETSI, 2014. European telecommunications standards

institute Members, http://www.etsi.org/index.php

/membership.

Anderson E., Bai Z., Bishof C., Demmel J., Dongarra J.,

1999. Lapack User Guide. Third Edition.

http://www.netlib.org/lapack/lug/

CULA., 2014. CULA Programmer’s Guide.

http://www.culatools.com/

Tomov S., Nath R., Du P., Dongarra J., 2011 MAGMA

Users’ Guide. http://icl.cs.utk.edu/projectsfiles/

magma/doxygen/

Agrell E., Eriksson T., Vardy A., Zeger K., 2002. Closest

point search in lattices. IEEE Transactions on

Information Theory, 48(8):2201-2214.

Schnorr C., Euchner M., 1994. Lattice basis reduction:

Improved practical algorithms and solving subset sum

problems. Mathematical Programming, 66(2):181-191.

Berenguer I., Wang X., 2003. Space-Time coding and

signal processing for MIMO communications. Journal

of Computer Science and Technology, 18(6):689-702.

Barbero L.G., Thompson J.S., 2008. Fixing the complexity

of the sphere decoder for MIMO detection, IEEE

Transactions on Wireless Communications, 7(6):

2131-2142.

Guo Z., Nilsson P., 2006. Algorithm and implementation

of the K-Best Sphere Decoding for MIMO Detection.

IEEE Journal on Selected Areas in Communications,

24(3):491-503., March 2006.

Su K., 2005. Efficient Maximum Likelihood detection for

communication over MIMO channels. University of

Cambridge, Technical Report.

Caire G., Taricco G., Biglieri E., 1998. Bit-interleaved

coded modulation. IEEE Transactions on Information

Theory, 44(3):927-946.

Barbero L.G., Ratnarajah T., Cowan C., 2008. A low-

PECCS2015-DoctoralConsortium

complexity soft-MIMO detector based on the fixed-

complexity sphere decoder. IEEE International

Conference on Acoustics, Speech and Signal

Processing. Las Vegas, Nevada (USA).

Müller-Weinfurtner S.H., 2002. Coding approaches for

multiple antenna transmission in fast fading and

OFDM. IEEE Transactions on Signal Processing, 50:

2442-2450.

Windpassinger C., Fischer R., Vencel T., H., Huber, 2004.

Precoding in multiantenna and multiuser

communications. IEEE Transactions on

Communications, 3(4):2057-2060.

Yao H., Wornell G., 2002. Lattice-reduction-aided

detectors for MIMO communication systems. IEEE

Global Communications Conference, Taipei, Taiwan.

MIMOPack-AHPCLibraryforMIMOCommunicationSystems