From Simulation to Deep Learning: Survey on Network Performance Modeling Approaches

Carlos Güemes-Palau; Miquel Ferriol-Galmés; Jordi Paillisse-Vilanova; Pere Barlet-Ros; Albert Cabellos-Aparicio

cs.NIMar 30, 2026

From Simulation to Deep Learning: Survey on Network Performance Modeling Approaches

Authors: Carlos Güemes-Palau, Miquel Ferriol-Galmés, Jordi Paillisse-Vilanova, Pere Barlet-Ros, Albert Cabellos-Aparicio

Organizations: Universitat Politècnica de Catalunya (UPC), Barcelona, Catalonia, Spain

Abstract

Network performance modeling is a field that predates early computer networks and the beginning of the Internet. It aims to predict the traffic performance of packet flows in a given network. Its applications range from network planning and troubleshooting to feeding information to network controllers for configuration optimization. Traditional network performance modeling has relied heavily on Discrete Event Simulation (DES) and analytical methods grounded in mathematical theories such as Queuing Theory and Network Calculus. However, as of late, we have observed a paradigm shift, with attempts to obtain efficient Parallel DES, the surge of Machine Learning models, and their integration with other methodologies in hybrid approaches. This has resulted in a great variety of modeling approaches, each with its strengths and often tailored to specific scenarios or requirements. In this paper, we comprehensively survey the relevant network performance modeling approaches for wired networks over the last decades. With this understanding, we also define a taxonomy of approaches, summarizing our understanding of the state-of-the-art and how both technology and the concerns of the research community evolve over time. Finally, we also consider how these models are evaluated, how their different nature results in different evaluation requirements and goals, and how this may complicate their comparison.

Figures & tables

Figure 1: Identified models per year of publication and type. Does not consider other types of publications, such as surveys. If the same model is covered in multiple papers, it is counted once in the year of its earliest publication.

Acronym	Description
ADNC	Algebraic-based Deterministic Network Calculus
AI	Artificial Intelligence
AQM	Active Queue Management
CVAE	Conditional Variational Auto-Encoder
[P]DES	[Parallel] Discrete Event Simulation
DCQCN	Data Center Quantized Congestion Notification [ 151 ]

Table 2

RON	Resilient Overlay Networks [ 120 ]
RTT	Round Trip Time
SDE	(System of) Stochastic Differential Equations
SFA	Separate Flow Analysis [ 93 ]
SNC	Stochastic Network Calculus
SVR	Support Vector Regression
[DC]TCP	[Data Center [ 5 ] ] Transmission Control Protocol

Table 1: List of acronyms

Figure 2: Taxonomy of network performance models. Solid line arrows indicate direct evolution, while dashed line arrows indicate inspiration or a “response to". Year indicates the year of the earliest publication within that category.

Model	Type	Input scope	Output scope	Traffic type	Performance metrics
REAL [ 82 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information
ns [ 136 ] , -2, -3 [ 121 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information
OMNeT, OMNeT++ [ 137 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information
OPNET [ 27 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information
CaSiNo [ 61 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information
WNS [ 22 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information

Table 5

pd-gem5 [ 4 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information	Testbed data
SimBricks [ 100 ]	Modular DES	Full network scenarios	Packet-level	Any	Full packet path information	Simulated data
SplitSim [ 99 ]	Modular DES	Full network scenarios	Packet-level	Any	Full packet path information	Simulated data
DONS [ 50 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information	Simulated data
Unison [ 15 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information	Simulated data
NSX [ 83 ]	DES	Full network scenarios	Packet-level	Any	Full packet path information	Only evaluates cost, not accuracy
Parsimon [ 150 ]	Link-level DES	Full network scenarios	Flow-level	TCP (Focuses on DCTCP, but generalizable)	Tail FCT (90th percentile or higher)	Simulated data

Table 2: Summary of simulated network performance models

Model	Type	Input scope	Output scope	Traffic type	Performance metrics	Evaluation
[ 88 ]	M/M/1, M/G/ $\infty$ , M/G/1 queues	Single-flow scenarios	Flow-level	Non-specific	Average packet delay, buffer overflow probability	Analytical evaluation
[ 117 ]	M/G/ $\infty$ queue	Traffic-matrix scenarios	Flow-level	Non-specific	Loss rate	Simulated data evaluation
[ 46 ]	GI/G/1 queue	Single-flow scenarios	Flow-level	Non-specific	Upper and lower bounds, average packet delay	Simulated data evaluation
[ 51 ]	System of M/M/ $\infty$ queues	Traffic-matrix scenarios	Flow-level	TCP (Tahoe)	Average packet [loss probability, RTT]	Simulated data evaluation
[ 52 ]	System of M/G/ $\infty$ queues	Traffic-matrix scenarios	Flow-level	TCP (Tahoe)	Average packet [loss probability, RTT]	Simulated data evaluation
[ 147 ]	Modified M/M/1 queues	Traffic-matrix scenarios	Flow-level	Non-specific	Flow throughput, average packet delay	Simulated data evaluation

Table 7

[ 109 ]	SDE fluid queuing model	Full network scenarios	Flow-level (temporal)	TCP (Generic w/ RED)	Flow throughput, average RTT (but evaluation only evaluates queue-length)	Simulated data evaluation
[ 103 ]	SDE fluid queuing model	Full network scenarios	Flow-level (temporal)	TCP (Sack, Reno, and NewReno)	Flow throughput, average RTT (but evaluation only evaluates queue-length and window size)	Simulated data evaluation
multi-AIMD [ 13 ]	Fluid/ discrete hybrid queuing model	Full network scenarios	Flow-level (temporal)	TCP and UDP	Flow throughput, average RTT	Simulated data evaluation
[ 18 ] , [ 94 ]	Fluid/ discrete hybrid queuing model	Full network scenarios	Flow-level (temporal)	TCP and UDP	Average flow throughput and RTT (only drop rate and window size retain a temporal component)	Simulated data evaluation
[ 105 ]	PDE fluid queuing model	Traffic-matrix scenarios	Flow-level (temporal)	TCP (Generic)	Average flow throughput, FCT and loss rate	Simulated data evaluation
[ 11 ]	PDE fluid queuing model	Traffic-matrix scenarios	Flow-level (temporal)	TCP (Scalable TCP)	Flow throughput	Simulated data evaluation
[ 23 ]	ODE fluid queuing model	Full network scenarios	Flow-level (temporal)	UDP 2 and TCP (Reno)	Flow RTT, throughput and loss rate	Simulated data evaluation

Table 8

[ 35 ]	Discrete and fluid queuing model	Traffic-matrix scenarios	Flow-level (temporal)	TCP	Flow packet delay (evaluation focuses in average queue lengths)	Simulated data evaluation
TFA [ 33 , 34 ]	ADNC	Full network scenarios	Flow-level (temporal)	Non-specific	Worst-bound packet delay	Analytical evaluation
D-BIND [ 87 ]	ADNC	Full network scenarios	Flow-level (temporal)	Non-specific	Worst-bound packet delay	Real data evaluation
[ 2 ]	ADNC	Full network scenarios	Flow-level (temporal)	UDP and traffic with throughput guarantees	Worst-bound packet delay	Analytical evaluation
SFA [ 93 ]	ADNC	Full network scenarios	Flow-level (temporal)	Non-specific	Worst-bound packet delay	Analytical evaluation
PMOO [ 44 ]	ADNC	Full network scenarios	Flow-level (temporal)	Non-specific	Worst-bound packet delay	Simulated data evaluation
[ 92 ]	ADNC	Full network scenarios	Flow-level (temporal)	Non-specific	Worst-bound packet delay	Analytical evaluation

Table 9

[ 84 ]	ODNC	Full network scenarios	Flow-level (temporal)	Non-specific	Worst-bound packet delay	Simulated data evaluation
[ 29 ]	SNC	Traffic-matrix scenarios	Flow-level (temporal)	Non-specific	Stochastic worst-bound packet delay	Analytical evaluation
[ 79 ]	SNC	Traffic-matrix scenarios	Flow-level (temporal)	Non-specific	Stochastic worst-bound packet delay	Analytical evaluation
[ 7 ]	SNC	Traffic-matrix scenarios	Flow-level (temporal)	Non-specific	Stochastic worst-bound packet delay	Analytical evaluation
[ 91 ]	Analytical model from RFC definitions	Single-flow scenarios	Flow-level (temporal)	TCP (generic with phase effects and RED)	Flow throughput	Simulated data evaluation
[ 113 ]	Stochastic analytical model from observations	Single-flow scenarios	Flow-level (temporal)	TCP (Reno)	Flow packet rate and throughput	Real data evaluation
[ 106 ]	Utility maximization model	Traffic-matrix scenarios	Flow-level	TCP	Flow throughput	Analytical evaluation

Table 3: Summary of analytical network performance models

Model	Type	Input scope	Output scope	Traffic type	Performance metrics	Evaluation
[ 67 ]	MA, EWMA, Holt-Winters	Single-flow scenarios	Flow-level (temporal)	TCP (generic)	Flow throughput	Testbed data
PathPerf [ 108 ]	SVR	Single-flow scenarios	Flow-level (temporal)	TCP (generic)	Flow throughput	Tesbed data and real data
WISE [ 135 ]	Causal graph	Single-flow scenarios	Flow-level	HTTP connections	Network response time	Real data
[ 69 ]	EVT	Multi-flow scenarios	Flow-level	UDP	50th, 75th, 90th, 99th, 99.9th, 99.99th, 99.999th and 100th percentiles for packet delay and jitter	Testbed data
CLAAP [ 110 ]	RBFNN	Single-flow scenarios	Flow-level (temporal)	UDP	Flow latency	Real data
[ 107 ]	DNN	Multi-flow scenarios	Flow-level	UDP	Average packet delay	Simulated data

Table 11

RouteNet-Erlang [ 42 ]	Custom-designed MPNN	Full network scenarios	Flow-level	UDP	Average packet delay and jitter, loss rate	Simulated data
RouteNet-Fermi [ 41 ]	Custom-designed MPNN	Full network scenarios	Flow-level	UDP	Average packet delay and jitter, loss rate	Simulated data and testbed data
RouteNet-Gauss [ 60 ]	Custom-designed MPNN with temporal component	Full network scenarios	Flow-level (temporal)	UDP	[Average, median, and 90th, 95th, and 99th percentile] packet [delay, jitter]	Testbed data
[ 37 ]	Custom-designed MPNN with graph attention	Full network scenarios	Flow-level	UDP	Average packet delay	Simulated data
[ 32 ]	Custom-designed MPNN with graph attention	Full network scenarios	Flow-level	UDP	Average packet delay	Simulated data
[ 80 ]	Custom-designed MPNN	Full network scenarios	Flow-level	UDP	Average packet delay	Testbed data
[ 58 ]	Custom-designed MPNN	Full network scenarios	Flow-level	UDP	Average packet delay	Testbed data

Table 12

DeepComNet [ 56 ]	GGNN [ 101 ]	Full network scenarios	Flow-level	UDP and TCP	TCP flow throughput, UDP average packet delay	Simulated data
[ 75 ]	GGNN [ 101 ]	Full network scenarios	Flow-level	TCP (Reno, Cubic, Bic, Illinois, Veno, Vegas and Ledbat)	Flow RTT and throughput	Simulated data
[ 134 ]	GCN [ 86 ]	Full network scenarios	Flow-level	UDP and TCP	Average packet delay	Simulated data
EAGLE [ 102 ]	Custom-designed MPNN	Full network scenarios	Flow-level	UDP and TCP	Average packet delay and loss rate	Simulated data
xNet [ 71 ]	Custom-designed MPNN with temporal component	Full network scenarios	Flow-level (temporal)	UDP and TCP (DTCP)	Average packet delay, flow throughput and FCT	Simulated data
GLANCE [ 96 ]	Custom-designed MPNN	Full network scenarios	Flow-level	UDP	Average packet delay and jitter, flow throughput, loss rate	Simulated data
FlowSeer [ 39 ]	Ensemble of GCN [ 86 ] , DNN with attention and DNN	Full network scenarios	Flow-level	UDP	Average packet delay	Simulated data

Table 13

xWeaver [ 141 ]	CNN	Full network scenarios	Flow-level	UDP	Flow FCT and throughput	Simulated data and testbed data
Deep-Q [ 143 ]	CVAE [ 85 ] + LSTM [ 70 ]	Multi-flow scenarios	Flow-level (temporal)	UDP	Packet delay and loss rates	Testbed data
[ 38 ]	Transformer [ 138 ]	Single-flow scenarios	Packet-level	UDP	Packet delay	Simulated data
DeepQueueNet [ 145 ]	Transformer [ 138 ] + BiLSTM	Full network scenarios	Packet-level	UDP	Full packet path information	Simulated data
[ 64 , 65 ]	NP [ 53 ]	Multi-flow scenarios	Flow-level	Non-specific	Average packet delay, flow throughput and loss rates	Testbed data

Table 4: Summary of ML network performance models

Model	Main Type	Secondary Type	Input Scope	Output Scope	Traffic Type	Performance Metrics	Evaluation
Pantheon [ 144 ]	Network emulator (Mahimahi [ 112 ] )	Bayesian optimization	Full network scenarios	Packet-level	Any	Full packet path information	Real data
iBox [ 10 ]	Network emulator with a queue model	Bayesian optimization	Full network scenarios	Packet-level	Any	Full packet path information	Real data
Prophet [ 148 ]	NUM analytical model [ 81 ]	Gradient descent	Full network scenarios	Flow-level	TCP 1	Flow throughput	Simulated data
DeepTMA [ 54 , 55 ]	NC (TMA [ 20 ] )	GGNN [ 101 ]	Full network scenarios	Flow-level (temporal)	Non-specific	Worst-bound packet delay	Simulated data
[ 68 ]	GraphSage [ 62 ]	NC (TFA [ 33 , 34 ] , SFA [ 93 ] )	Full network scenarios	Flow-level	UDP	Mean, min, max, 90th and 99th percentile packet delays	Testbed data
QT-RouteNet [ 36 ]	RouteNet [ 123 ]	M/M/1/B queue model	Full network scenarios	Flow-level	UDP	Average packet delay	Simulated data

Table 15

QINN [ 66 ]	DNN	M/G/1 queue model	Traffic-matrix scenarios	Flow-level	UDP and TCP (non-specific)	Avg. packet delay, throughput	Simulated data
[ 146 ]	ns-2 [ 136 ]	Flow queueing model [ 103 ]	Full network scenarios	Packet-level	UDP and TCP 2	Full packet path information	Simulated data
MimicNet [ 149 ]	ns-3 [ 121 ]	LSTM [ 70 ]	Full network scenarios	Packet-level	Any	Full packet path information	Simulated data
m3 [ 97 ]	flowSim [ 111 ]	Transformer (LLama2 [ 73 ] )	Full network scenarios	Flow-level	UDP and TCP (DCTCP, TIMELY, DCQCN, HPCC)	99th percentile FCT	Simulated + Real data
CausalSim [ 6 ]	Trace simulation	Causal DNN	Full network scenarios	Packet-level	Any	Full packet path information	Simulated data
Sim2HW [ 130 ]	OMNET++ [ 137 ]	GraphSage [ 62 ]	Full network scenarios	Flow-level	UDP	Min, 25th, 50th, 75th, 90th, 99th, 99.9th, 99.99th, 99.999th percentile packet delays	Testbed data

Table 5: Summary of hybrid network performance models

Approach		Accuracy	Expressiveness	Applicability	Computational Cost
Simulation	DES	High	Packet-level	General	High (and sequential)
Simulation	PDES	High			High
QT	Discrete	Low	Flow-level	Specific TCP version	Low
	Fluid		Flow-level
			(temporal)
NC	ADNC	Medium	Flow-level	Feed-foward networks	Medium

Table 6: Summary of current network performance modeling approaches.

Figure 3: Evaluation categories across different model types. If a model is evaluated with multiple categories, they are categorized in the following priority: real data, testbed data, simulated data, and analytical.

Queueing Theory

See all 18 papers

Network-Aware Forecasting on Wireless Access Points
Sep 2, 2026Niloo Bahadori, Swadhin Pradhan, Peiman AminiEfficient Neural Network Inference Edge Computing
Cisco Systems
Network World Models as Environments for Algorithm Design on Complex Systems
Oct 1, 2026Rishab Alagharu, Hongji Pu, Zeeshan Memon +3Automated Algorithm Discovery World Model-Based Planning
Emory University · University of Illinois Urbana-Champaign
Telecom World Models: Unifying Digital Twins, Foundation Models, and Predictive Planning for 6G
Apr 8, 2026Hang Zou, Yuzhi Yang, Lina Bariah +15Network Slicing World Model-Based Planning
Research Institute for Digital Future, Khalifa University, 127788 Abu Dhabi, UAE · College of Information Science and Electronic Engineering, Zhejiang University, 310027, Hangzhou, China · University of Oulu and Aalto University +2

Sep 2, 2026J/K move · Enter open · S save

Network-Aware Forecasting on Wireless Access Points

Niloo Bahadori, Swadhin Pradhan, Peiman Amini

Cisco Systems

Efficient Neural Network Inference Edge Computing Edge ML Wireless Local Area Networks

Open paper AskPDF arXiv

Enterprise wireless access points (APs) are promising platforms for predictive machine learning (ML), but their primary responsibility remains providing wireless connectivity and network services. Predictive inference must therefore share an AP's CPU and memory with packet processing, Wi-Fi and IoT radio operations, and client management. This resource contention creates two risks: a model that performs well on proxy hardware may be too slow on the target AP, while a model that fits in isolation may still degrade network services under load. We define network-aware deployability using two gates: qualification of the model and its execution path on the target AP, followed by validation of its execution profile under packet-service and forecasting constraints. Our benchmarks show that edge testbeds do not reliably capture target behavior. Across matched artifacts and serving settings, five model implementations run 6.1--19.1

\times

slower on an AP than on a Raspberry Pi5, while peak memory usage differs by up to 22%. Moreover, two forecasting foundation models of similar size differ in AP latency by 19 $\times$ . When serving a smaller model across 13 parallel streams at a 30s cadence under network saturation, default execution increases p99 round-trip time (RTT) by 76% and reduces throughput by 7.06%. Understanding these trade-offs is essential for live deployment if we aim to use APs for both networking and ML workloads.

Abstract

Queueing Theory

Explore similar work

Network-Aware Forecasting on Wireless Access Points

Network World Models as Environments for Algorithm Design on Complex Systems

Telecom World Models: Unifying Digital Twins, Foundation Models, and Predictive Planning for 6G