Wireless Information Network Laboratory (WINLAB). Rutgers University, 73 Brett Road. Piscataway, NJ ... been studied at Georgia Tech and Bellcore [16, 17, 18, 19] .... The base zone entity contains TED processes for MS admission, for MS ...
Copyright 1998 IEEE. Published in the Proceedings of PADS’98, May 26-29, 1998 in Banff, Alberta, Canada. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE. Contact: Manager, Copyrights and Permissions / IEEE Service Center / 445 Hoes Lane / P.O. Box 1331 / Piscataway, NJ 08855-1331, USA. Telephone: + Intl. 908-562-3966.
W I PPET, A VIRTUAL TESTBED FOR PARALLEL SIMULATIONS OF WIRELESS NETWORKS Jignesh Panchal, Owen Kelly, Jie Lai, Narayan Mandayam, Andrew T. Ogielski, Roy Yates Wireless Information Network Laboratory (WINLAB) Rutgers University, 73 Brett Road Piscataway, NJ 08854–8060, U.S.A. Abstract We describe the T E D/C++ implementation of W I PPET, a parallel simulation testbed for evaluating radio resource management algorithms and wireless transport protocols. Versions 0.3 and 0.4 of the testbed model radio propagation (long– and short–scale fading and interference) and protocols for integrated radio resource management in mobile wireless voice networks including the standards– based AMPS, NA–TDMA and GSM protocols, and several research–oriented protocol families. We provide parallel performance data verifying that the dominant computational demand due to received signal quality calculation can be partitioned geographically, by orthogonal radio channels, or in a hybrid manner. 1 Introduction We begin with a discussion of the simulator itself, its purpose, and the nature of the phenomena it models. Simulations of wireless networks differ considerably from simulations of wired networks in that they have to include modeling of the geographic environment and physical channel properties (fading and interference) which both have a profound effect on network performance. Frequent calculations of received signal quality for each network node are necessary to emulate node behavior in the presence of fluctuating signals and these calculations dominate simulation. W I PPET1 is a versatile simulator for wireless networks, which has been implemented in the T E D/C++ framework [1, 2, 3, 4], and is used with the Georgia Time Warp (GTW) optimistic parallel simulator [5]. W I PPET is a completely redesigned and substantially extended successor to WINLAB’s older, sequential simulator, MADRAS [6], the Mobility And Dynamic Resource Allocation Simulator. MADRAS has figured importantly in research (e.g., [7, 8, 9, 10]), but its monolithic design has 1 Wireless Propagation and Protocol Evaluation Testbed. Also a breed of fast racing dogs (whippet).
led to demands for an object–oriented, reusable testbed approach, as well as for faster execution. W I PPET is a collection of base T E D objects modeling mobility, traffic radio propagation and protocols that together with a configuration file (specifying interconnections and system parameters) can be instantiated as a virtual testbed [11]. The objectives of the testbed are to: (i) Shift the work to design of models rather than low–level programming of custom simulators, and (ii) Enable reuse of a growing database of objects for creation of new simulation models. Therefore, the design of base entities requires the efficient implementation of the fundamental behaviors, while expressing the unifying features of the different design options to enable entity substitutions. The paper [11] provides more discussion on casting the wireless network into reusable and substitutable parts; in the current work we emphasize efficiency. Because of the dominance of the interference calculation, the pursuit of efficiency leads always back to the zone entity which calculates interference (using long– and short–scale fading models) at granularities appropriate to studies at differing levels of the network hierarchy. This article describes and evaluates the alpha versions W I PPET 0.3 and W I PPET 0.4 which implement differing zone designs in a cellular voice network simulation. Current and near-term research involving simulations with W I PPET focuses on radio resource management and media access protocols, effects of packet losses due to interference and fading on the performance of Internet protocols, and studies of interference among nearby autonomous systems employing distinct modulation and radio resource management algorithms, especially in the 5 GHz U-NII bands. An important application area is student education and training in wireless network principles. In related work, parallel discrete event simulation of mobile wireless networks have been investigated by few groups. Ad-hoc multi-hop radio networks have been investigated at UCLA [12, 13, 14] incorporating the SIRCIM [15] indoor fading simulator. PCS systems have been studied at Georgia Tech and Bellcore [16, 17, 18, 19]
and at KTH [20]. There are many papers on sequential simulation of wireless networks and of radio channels, but their review is beyond the scope of this article. 2
W I PPET design
Wireless network simulation is fundamentally about assessing the interaction of mobility, traffic, propagation, and protocols (radio resource management). In this section we present background material on these phenomena leading to the system partitions shown in Figure 2 (external view) and Figure 3 (internal view). We discuss the basic design, mobility, and traffic in this introduction and devote separate subsections to propagation and radio resource management. The section concludes with a description of W I PPET 0.3/0.4 functionality. The design of primary T E D entities for a wireless network simulation has to be convenient for the modelers. Natural candidates for entities are mobile stations, base stations and switches. Thus entities MS, BS, and switch each represent their physical counterpart. These entities contain T E D processes simulating radio resource management protocols at the connection level (session admission, transmitter power control and handoff). There are two additional entities, callgen which generates calls (session arrivals), and zone which is described in Section 2.2. A mobile wireless simulator must include a description of the terrain in which the mobile stations (MS) can travel. Moreover, the geometry of the terrain determines gross radio propagation characteristics. In W I PPET, the continuous terrain is approximated by an adjacency graph of accessible MS locations. Mobility describes the process by which MS move from one point to another within the modeled geography. W I PPET allows importation of an approximation to a real map, or use a synthetic geography. Similarly it may be driven by precomputed mobile trajectories (a particular traffic scenario) or have trajectories that evolve at runtime. The simulator provides for the generation of offered load (traffic) to the network through the callgen entity. For session–level simulations the parameters of interest include the rate of call arrival and spatial and temporal variations in that rate, and the parameters of the individual call. The life cycle of a call begins at the callgen entity where the arrival time, initial location, duration, and radio resource requirements are selected and loaded into a blank MS entity. At the arrival time, the MS awakes, begins moving, and notifies the zone about its arrival and to provide necessary radio resources. In the current version, every user requests a voice channel; calls arrive according to a Poisson process, distributed uniformly in space; and calls have exponentially distributed holding times. New traffic and user classes are added by defining new architectures
for the callgen, MS and BS entities. 2.1 Propagation and interference Performance of a radio receiver depends on the quality of the received signal from the selected transmitter. Received signal quality exhibits very large fluctuations in space and time (fading) due to environmental motion and activity of other sources, and their modeling is essential for wireless network simulators. There is a vast body of literature on the measurements and modeling of propagation in various environments at various frequencies, for a review see [21, 22]. The classical models are derived as follows: The power received at location x from a single transmitter at b over a general time-varying linear radio channel is PRX (t; x; b) =
Z
t
?1
h(t; ; x; b)PTX (; b) d
(1)
where h is the impulse response. A quantity of significant interest is signal to interference ratio (SIR), which for the transmission from b to x is defined by the ratio PRX (t; x; b)=I (t; x; b), where I (t; x; b) is a sum of contributions of all other transmitters sharing the spectrum plus an additive noise term and is given as I (t; x; b) =
X
b 6=b
PRX (t; x; b0 ) + N0
:
(2)
0
The modeling of the channel response h(t; ; x; b) depends on performance assessment needs: it may be site– specific, that is based on actual measurements or on raytracing approximation for a particular town or building; or it may employ stochastic processes with values and correlations mimicking those measured in a class of of similar environments (e.g., typical American suburbs, or generic multi-story office buildings). While the design of W I PPET allows use of site–specific propagation models, our research focuses on statistical characterizations of radio networks, and thus we limit the discussion to stochastic models of h(). The impulse response of the radio channel is commonly represented as the product of long-scale and short-scale fading; i.e. h(t; ; x; b) = g (x; b)h(t; ). The long-scale fading, g (; ), represents the attenuation of the signal averaged in a region larger than 10 or more carrier wavelengths around the receiver, thus due to distance and shadowing by large objects. The short-scale factor h(t; ) represents time-varying effects of interference and delay dispersion of carrier waves arriving over distinct paths. It is important to note that the details of the fading model should be matched to the signal processing capabilities of the modeled receiver to obtain the meaningful statistics of
data loss; i.e., further abstractions of fading may be made according to the level in Figure 1 (top) that is of interest. For example, studies of radio resource management typically require averaged fading and interference, whereas for packet-level simulations one needs SIR values on a fine time scale during a packet to determine if the packet’s data have been corrupted. Sum (2) results in a cross–connection of transmitters to receivers as illustrated in Figure 1 (bottom). However, a transmitter and receiver may act independently if they do not share the same spectrum, or if they are sufficiently separated in distance that g (x; b) 0. That is, depending on the structure of system being simulated, the cross–connection may be essentially complete, or relatively sparse.
ing contributions to total signal power in a radio channel from all active transmitters and attenuating them by appropriate long- and short-scale fading factors. To facilitate mobility and received power calculations, the zone entity contains: (i) a large constant structure containing the geography points together with precomputed path loss matrix g (; ); (ii) for each radio channel, a dynamic list of data structures characterizing the stations transmitting in this radio channel. For example, for each MS such a data structure contains instantaneous values of position, transmit power, assigned base station ID, etc.; (iii) for each transmitter-receiver pair, a dynamic list of data structures characterizing current state of short-scale fading. The base zone entity contains T E D processes for MS admission, for MS position updates (mobility), for short-scale fading generation, and for reporting the signal power to all active receivers. Note that a realistic simulation like the one partially described here necessarily contains a very large state, which adversely affects simulation speed of optimistic parallel simulators during state-saving and rollbacks. 2.3 Two parallelizations of zone
Figure 1: Top: Abstraction of the generic communicating packet radio entities. Bottom: The main role of a zone entity is to combine the contributions of all actively transmitting Tx to the received power profile for every Rx. Each time the power, or position, or the value of short–scale fading loss of a Tx changes, all active Rx power profiles in that radio channel must be updated.
2.2 Zone design Encapsulation of geography and propagation in T E D entities in a way that promotes parallel simulation efficiency is not obvious. What is clear is that over a wide range of temporal scales, geography, mobility, and propagation remain strongly coupled in the calculation of received power, through equation (2). Because the simulations are dominated by frequent calculations of received signal quality, our goal is to identify canonical patterns of partitioning received power calculations to T E D entities which can be executed in parallel. We therefore choose to encapsulate these coupled quantities jointly in an entity named zone whose structure and role is sketched below. A zone entity combines the role of a geography container with the role of a radio channel simulator, collect-
It is advantageous to divide and conquer the zone’s heavy computational demand by partitioning the model, when possible, into multiple zones for parallel simulation. Figure 2 shows two canonical ways that a zone may be subdivided: (i) By geographic parallelism, wherein a large model geography is partitioned into zones, with each zone containing a fraction of the geography points together with corresponding path loss submatrix; (ii) By radio channel parallelism for multi-channel radio networks, wherein each zone contains all geography and path loss data, but only a fraction of the radio channels. Note that mobiles may leave or enter the zone, due to motion in case (i) or due to channel reassignment in case (ii). Then, a zone transfer is executed by which state information regarding the exiting mobile is transferred from the old zone to the new zone. Since a zone outweighs any other entity, having a partitionable zone entity facilitates load balancing (see Section 3). 2.4 Wireless protocols and entity hierarchy The fourth of our four essential elements of wireless network simulation is protocols, including radio resource management. Radio resource management includes any actions that are necessary in a wireless system to provide the required Quality of Service (QoS). In wireless voice networks the term has generally included protocols that control call admission, channel allocation, power control, and handoff. In a wireless data network there are also transport protocols acting above the resource management activities.
essentially a distillation of the unifying concepts underlying the task that the protocol performs. For example, W I PPET implements four standards-based handoff algorithms for cellular telephony, plus three others from the literature. Although these protocols had as many as fourteen defined messages, careful comparison revealed that all of the messages over all of the protocols carry only three distinct payload types! Thus three event types comprise the W I PPET handoff event set, and each carries a field denoting which protocol variant it is acting for, and within that, which of the original messages it represents. Further, because different instances of an entity are able to bind different architectures, it is possible simulate a nonhomogeneous population of users: GSM phones together with AMPS phones for example.
call generator
MS
MS
MS
MS
MS
switch
BS BS zone
zone
zone
zone
BS
channel channel
BS
MS
channel channel
BS
MS
MS
call generator
switch
BS
MS
Figure 2: Two approaches to parallel simulation of mobile, multi-channel wireless networks: either by partitioning the geographical terrain among zone entities (top, W I PPET 0.3), or by partitioning the spectrum into orthogonal radio channels among zone entities (bottom, W I PPET 0.4). Hybrid approaches are also possible. Lines between entities denote the principal event flows.
W I PPET’s primary purpose is evaluation of protocols, and evaluation means performance comparison. Within T E D, comparison is facilitated by the ability to substitute architectures of entities. A canonical comparative experiment requires running two nearly identical models, the only difference being that certain protocol processes are different. Unlike propagation and interference, which require considerable thought to cast into the event-passing paradigm of T E D, protocols fall easily into place by defining an event for each message type of the protocol, and a T E D process for each protocol, to evaluate these events. Our comparative experiments introduce the requirement that competing protocol variants must be able to bind to the same entity, and that corresponding protocol processes in different entity types must accept the same events for compatibility. The primary entities in W I PPET and their inheritance hierarchy are shown in Figure 3 where horizontal arrows indicate a compatibility requirement. The entity and event definitions should accommodate as many future protocol variants as the designer can envision:
Figure 3: Entity inheritance hierarchy (vertical) and their compatibility for protocol substitutions (horizontal). We note that in the current version of T E D it is computationally expensive to embed component entities in parent entities when they need to send or receive events through the parent’s channels. Therefore, our design builds up each entity type by inheritance of architectures, with each level of inheritance adding appropriate processes and state fragments.
2.5
W I PPET 0.3/0.4 functionality
Versions 0.3 and 0.4 of W I PPET implement simulations of a wireless cellular voice network. W I PPET 0.3 implements geographic zone partitioning and W I PPET 0.4 implements channel zone partitioning, but otherwise the two versions are functionally identical. For call admission W I PPET 0.3/0.4 has two variants of base station and channel assignment: use a channel from the base with the strongest pilot tone, or search a dynamic list of bases in order of decreasing pilot strength to find an acceptable channel. From [6] we have incorporated a family of dynamic channel allocation methods that employ interference measurements. Six handoff algorithms have been implemented. Three are based on standards: AMPS, North American TDMA (IS-136), and GSM. The other three algorithms, SIR based handoff, received-signal-strength (RSSI) based handoff, and a combined SIR/RSSI based handoff, are derived from [8]. Figure 4 shows an interaction diagram for AMPS handoff that exemplifies the
level of implementation detail. The power control algo-
same T E D/GTW simulator both for the one- and multiprocessor simulations. Table 1 specifies the Poisson call
AMPS handoff - interaction diagram
time
SWITCH
Parent BS
Candidates BSs
All ZONES
parent BS measurement request
handoff check
MS
calculate signal & interference power for MS
measurement handoff request find candidate BSs send requests for measurement
select target BS, handoff command to parent BS
measurement
handoff command
check for handoff condition measurement request invoke DCA, find channel, req. MS Tx pwrs.
disconnect traffic channel
update MS record
handoff command update MS state
handoff ACK
A 2 10 – 25
B 6 10 – 25
C 2 2 – 25
scenario D E 2 6 10 10 0.5 0.5 25 25
F 2 10 0.1 25
G 2 10 0.5 2.2
Table 1: Seven experimental scenarios to evaluate geographic and channel zone partitioning.
invoke DCA, find channel, req. MS Tx pwrs.
handoff command
parameter [units] traffic [calls/s] resolution [m] PWR period [s] speed [m/s]
handoff ACK
connect traffic channel handoff check
Figure 4: Interaction diagram for AMPS handoff. rithm is an asynchronous distributed uplink power control in which each user attempts to minimize its transmit power while maintaining a required SIR [23]. Long– scale fading is precomputed for the specified geography by overlaying correlated log–normal shadow fading [24] on power–law distance loss [22]. We have implemented the models of short–scale Rayleigh and Ricean fading [25] with (i) wave-superposition method, (ii) spectrum-shaping method; and also (iii) a quantized channel state Markovian model [26], but these are not employed in the experiment reported in Section 4. 3 Experiment The purpose of our experiment is to evaluate the parallel performance of geographic and channel zone partitioning and the sensitivity of that performance to simulation parameters. The geography is a synthetic Manhattan-style urban environment of size 12 12 km (24 24 city blocks). Mobile trajectories are modeled by a biased random walk. Radio propagation is modeled with a spatial stochastic process of correlated log-normal fluctuations superimposed on distance dependent path loss factors characteristic for this category of urban environments. The study area contains 48 base stations sharing 40 radio channels with reuse factor one. Call admission and handoff are based on RSSI; power control is based on SIR. Call holding time is exponentially distributed with mean 120 seconds. Total simulated time in each case has been 10,000 seconds. For speed comparisons we used the
arrival rate, the geographic resolution, the signal checking period, and the travel speed for seven scenarios that were tested with geographic and channel partitioning on up to 8 processors of a multiprocessor Sun Enterprise 4000 server. Signal quality checks for power adjustment and potentially to initiate handoff are executed asynchronously and independently for each MS with the PWR period specified in Table 1. Scenarios A, B, and C simulate only mobility and call admission. A is the nominal case among the mobility scenarios, B has 3 offered traffic, and C has 5 finer geographic resolution. Scenarios D, E, F, and G simulate mobility, call admission, power control and handoff. D is the nominal case among the interference scenarios, E has 3 offered traffic, F has 5 more frequent power adjustments, and G has 11 slower moving subscribers. In the nominal interference scenario (D), an average of 240 calls are active at any moment. Model partitioning depends on the number of processors used. The mapping strategy has the number of processors equal to the number of zones, with one zone mapped to each processor. An equal number of mobiles are mapped to each processor, and similarly for bases. Callgen and switch, of which there are one each, are mapped to different processors. 4 Parallel performance analysis After all these preparations, the big question is: How much faster can we run the simulations? Results appear in Figures 5–8. Figures 5–7 share a similar two-by-two panel layout such that the left column refers to geographic zone partitioning, the right column refers to channel zone partitioning, the top row refers to mobility scenarios (A, B, C), and the bottom row refers to interference scenarios (D, E, F, G). The results in Figures 5 show a strong speedup in some but not all of the scenarios. It is interesting to analyze how this happens. First, there is a considerable concurrency in the system itself — a prerequisite for any speedup at all. From the simulation standpoint, however, that is not
geographic partition
channel partition
6
6 A: mobility B: 3× traffic C: 5× resolution
speedup
5
5
4
4
3
3
2
2
1
1 1
2
4
6
8
6
2
1
2
4
6
8
6 D: interference E: 3× traffic F: 5× PWR updates G: 11× slower
5 speedup
1
4
5 4
3
3
2
2
1
1 1
2
4 6 number of processors
8
N
4 6 number of processors
8
Figure 5: The speedup for processors is defined as the ratio of computing time on one processor to the computing time on processors.
N
enough; we must also consider the costs of partitioning the problem into communicating submodels, and the mechanisms that force synchronization between submodels. It is helpful to think of each scenario as a blend of pure mobility activities and pure interference calculation activities. The next four paragraphs consider partitioning costs and synchronization mechanisms for the four combinations of activity and partitioning method. Mobility/geographic: Mobility is an autonomous activity of each mobile that is managed by its zone. Partitioning introduces zone transfer events. Synchronization is forced when an MS transfers from one zone to another by moving. Mobility/channel: In channel partitioning, zone transfers are also introduced and synchronization is forced when an MS transfers from one zone to another by changing channel. In both channel and geographic partitioning, zone transfers happens at most a handful of times per call. The low cost and rarity of synchronization required for pure mobility is supported by the strong speedup curves in the top panels of Figures 5. In those panels, note especially scenario C which has the largest fraction of mobility effort among the seven scenarios. Interference/channel: For orthogonal radio channels, interference in one channel is completely independent of interference on other channels. Zone transfer is the only event introduced by partitioning. Here again, synchronization is forced only when the MS transfers from one zone to another by changing channel. The rare synchronization required with orthogonal radio channels is supported by the strong speedup curves in the lower right panel of Figures 5. Interference/geographic: Within a radio channel, inter-
ference is a simultaneous phenomenon across the geographical extent of the network. Therefore the cost of geographic partitioning is that for each of the hundreds of interference requests that are processed during a single, a wave of request/response events carrying portions of the sum in equation (2) ripples between the current BS and the entire array of zones . The benefit is that the processing time for a portion of the sum is less than the processing time for the whole sum. All zones are globally synchronized at each interference request. Can we blame this extensive synchronization requirement for the dramatic failure of geographic parallelism seen in the lower left panel of Figures 5 (which shows virtually no speedup with increasing number of processors)? Surprisingly not. In fact the lower panels of Figure 6 show that the efficiency of the optimistic simulator (the ratio of committed events to total events) is nearly identical for interference scenarios under geographic versus channel partitioning. The culprit is the total number of events. We have touched on the fact that partitioning of a model into multiple zones necessarily introduces additional events, state variables and processes into the simulation. These additions do not change the behavior of the modeled system, but act to coordinate calculations involving two or more zones. Figure 6 quantifies the effects of partitioning on event counts. Total events processed is the sum of committed events and rolledback events. The number of committed events is a measure of the size of the specified simulation, whereas the number of rolledback events is a measure of the overhead incurred by use of optimistic simulation. The lower left panel shows that even without rollbacks, partitioning the interference sum necessitates a proportionate increase in messaging. That investment can only be recovered if either the amount of computing per synchronized signal quality calculation is much larger than event processing overhead, or if the geographical fan-out of messages is reduced. The signal quality calculations in the current experiment are lightweight with respect to the event processing overhead the current T E D/C++ . Geographical fan-out could be limited in a system where the study area was significantly larger than the effective radio range so that only a base-dependent subset of zones would need polling to fulfill each interference sum. Preliminary experiments in reducing event counts, and in artificially increasing the work of interference calculation show that both mechanisms serve to improve the speedup seen in geographical parallelization. There is also scope to reduce the parallelization costs by simple programming techniques that reduce event usage, such as event aggregation or subscriber notifications. When a system is partitioned along the spectral direction into zones containing a subgroup of radio channels each, asynchronous computations of mobility and interference both contribute towards high exploitation of inherent
geographic partition
channel partition 15
A: committed events rolledback events
10
10
5
5
0
0 1
2
4
6
8
2
4
6
8
1
2 4 6 8 number of processors
400 D: committed events rolledback events
300
300
200
200
100
100 0 1
2 4 6 8 number of processors
Figure 6: The figures show numbers of committed and rolledback events to perform the same 10000 second simulation for the nominal mobility scenario (A) and the nominal interference scenario (D). Subdividing zones requires additional events to communicate results among zones and for zone-transfer of mobiles. The effect is especially evident in geographically parallelized interference calculations.
concurrency that promotes fast parallel execution. However, channel parallelization does not accommodate systems that lack an orthogonal decomposition. In certain cases it may be best to have only one instance of a zone entity. In contrast, signal quality calculations in geographically partitioned zones force substantial communication between submodels that can worsen parallel performance considerably. It is reasonable to ask why one would simulate a system whose geographical dimension exceeded the radio range. It is true that for straightforward statistical characterizations, one would set the system size approximately equal to the effective radio range. However, W I PPET is specifically designed so that it may be subsumed into a larger simulation [1]. Such a simulation might be working to statistically characterize a higher level network phenomenon whose physical extent exceeds radio range, or it might be a predictive simulation over a real geography of large extent. In a similar vein, one might also ask why we partition the functionality of a wireless network rather than, say, a temporal partition (i.e. simulate 10000 seconds, by running 10 versions of a sequential simulator for 1000 seconds each). The answer again is that we might wish to investigate network phenomena that exhibit long-term correlations, or to drive simulations with lengthy traces from actual networks. These are not toy models, but elaborate standards-based simulations of realistic systems, and the results shown herein indicate that parallel simulation research ready for
geographic partition average RB distance [events]
0
channel partition
60
60 A: mobility B: 3× traffic C: 5× resolution
50
50
40
40
30
30
20
20
10
10
0
0 1
average RB distance [events]
millions of events
400
1
production use. Simulation time can be significantly reduced by parallel execution when the T E D model is large enough for a given number of processors, and its implementation follows a design pattern that exploits well the available concurrency. In summary, this and related work clearly indicates that once the mechanisms for parallelizing the simulations of wireless networks are implemented in the design of primary simulation objects (which is not easy), modelers can routinely reuse them to create models exhibiting significant speedups: even when only a 4 processor workstation is used, speeding the simulations by a factor of 3 means a lot in engineering design! Orthogonal channels and limited radio range are the primary opportunities for exploitable concurrency in wireless networks.
2
4
6
8
6
1
2
1
2
4
6
8
6 D: interference E: 3× traffic F: 5× PWR updates G: 11× slower
5 4
5 4
3
3
2
2
1
1
0
0 1
2
4 6 number of processors
8
4 6 number of processors
8
Figure 7: The figures shows the average length of rollbacks registered by the GTW optimistic parallel simulator. Scenarios that have less frequent synchronization exhibit larger rollback distances. (The net effect on efficiency depends also on the number of rollbacks.)
A
geographic partition code channel partition code
B scenario
millions of events
15
C D E F G 0
1000 2000 3000 4000 5000 6000 7000 8000 9000 sequential execution time for 10000s simulation time [seconds]
10000
Figure 8: Execution times that form the baseline of Figure 5 show the relative expense of the seven test scenarios. When only one zone is in use, the geographic and channel parallelizations perform the same calculation (i.e. they collect interference contributions from all channels over the whole study area); the difference in execution times seen in the figure is due to minor code differences between W I PPET versions 0.3 and 0.4.
Acknowledgments This work is partially supported by DARPA Contract N66001-96-C-8530 and by NSF Grant NCR-9527163. We thank Kalyan Perumalla for useful discussions of T E D/C++ , James Cowie for discussions on software frameworks and design patterns in system modeling, and Sorabh Gupta, Vipin Sali, and David Pandian for separate contributions. References
[12] Joel Short, Rajive Bagrodia, and Leonard Kleinrock. Mobile wireless network system simulation. Technical Report 950015, Computer Science Department, UCLA, Apr 1995. [13] R. Bagrodia, M. Gerla, L. Kleinrock, J. Short, and T-C. Tsai. Hierarchical simulation environment for mobile wireless networks. In Proceedings of the 1995 Winter Simulation Conference, WSC’95, pages 563–570, 1995. [14] W. Liu, C-C. Chiang, H-K. Wu, V. Jha, M. Gerla, and R. Bagrodia. Parallel simulation environment for mobile wireless networks. In Proceedings of the 1996 Winter Simulation Conference, WSC’96, pages 605–612, Coronado, CA, 1996.
[1] D. M. Nicol, editor. Special Issue on the Telecommunications Description Language, volume 25 no. 4 of Performance Evaluation Review. ACM SIGMETRICS, March 1998.
[15] Theodore S. Rappaport, Scott Y. Seidel, and Koichiro Takamizawa. Statistical channel impulse response models for factory and open plan building radio communication system design. IEEE Transactions on Communications, 39(5):794–807, May 1991.
[2] K. Perumalla, A. Ogielski, and R. Fujimoto. MetaTeD: A meta language for modeling telecommunication networks. TR GIT-CC-96-32, College of Computing, Georgia Institute of Technology, 1996.
[16] R. M. Fujimoto C. D. Carothers and Y.-B. Lin. A Case Study in Simulating PCS Networks Using Time Warp. Proceedings of the 9th Workshop on Parallel and Distributed Simulation, pages 87–94, 1995.
[3] K. Perumalla and R. Fujimoto. A C++ instance of TeD. TR GIT-CC-96-33, College of Computing, Georgia Institute of Technology, 1996.
[17] C. D. Carothers, Y.-B. Lin, and R. M. Fujimoto. A re-dial model for personal communications services networks. In 1995 Vehicular Technology Conference (VTC ’95), pages 135–139, July 1995.
[4] K. S. Perumalla, M. Andrews, and S. Bhatt. A virtual PNNI network testbed. In Winter Simulation Conference Proceedings, Dec 1997. [5] S. Das, R. M. Fujimoto, K. Panesar, D. Allison, and M. Hybinette. GTW: A time warp system for shared memory multiprocessors. In Winter Simulation Conference Proceedings, pages 1332–1339, 1994. [6] Kevin Gang Chen. Integrated dynamic radio resource management of wireless communication systems. Masters thesis, Rutgers, The State University of New Jersey, New Brunswick, NJ, May 1996. [7] C.N. Chuah and R. Yates. Evaluation of a minimum power handoff algorithm. In Sixth IEEE International Symposium on Personal, Indoor and Mobile Radio Communications PIMRC’95, pages 623–627, 1995. [8] C.N. Chuah, R. Yates, and D. Goodman. Integrated dynamic radio resource managemant. In Proc. IEEE Vehicular Technology Conference, VTC-95, pages 584–588, 1995. [9] R. Yates, S. Gupta, C. Rose, and S. Sohn. Soft dropping power control. In Proceedings of The IEEE Vehicular Technology Conference VTC 97, May 1997. [10] R. Yates and C.Y. Huang. Integrated power control and base station assignment. IEEE Transactions on Vehicular Technology, 44(3):638–644, August 1995. [11] J. Panchal, O. Kelly, J. Lai, N. Mandayam, A. T. Ogielski, and R. Yates. Parallel simulations of wireless networks with TED: Radio propagation, mobility and protocols. Performance Evaluation Review, 25(4):30–39, Mar 1998.
[18] C. D. Carothers, R. M. Fujimoto, Y-B. Lin, and P. England. Distributed simulation of large-scale pcs networks. In Proceedings of the 1994 MASCOTS Conference, pages 2–6, January 1994. [19] Y.-B. Lin and V. K. Mak. On Simulating a LargeScale Personal Communications Services Network. ACM TOMACS, 1993. [20] M. Liljenstam and R. Ayani. A model for parallel simulation of mobile telecommunication systems. In Proceedings of the International Workshop on Modelling, Analysis, and Simulation of Computer and Telecommunication Systems (MASCOTS), 1996. [21] W. C. Jakes, editor. Microwave Mobile Communications. Wiley, 1974. [22] Theodore S. Rappaport. Wireless Communications. Prentice Hall, 1996. [23] Roy D. Yates. A framework for uplink power control in cellular radio systems. IEEE J. Sel. Areas Commun., 13(7):1341–1346, Sep 1995. [24] M. Gudmundson. Correlation model for shadow fading in mobile radio systems. Electron. Lett., 27(23):2145–2146, 1991. [25] R. H. Clarke. A statistical theory of mobile-radio reception. Bell System Technical Journal, 47:957–1000, 1968. [26] F. Babich, O. E. Kelly, and G. Lombardi. A variable-order discrete model for the mobile propagation channel. In 9th Tyrrhenian International Workshop on Digital Communications, Lerici, Italy, September 7–10, 1997. IEEE.