Flexible Performance Prediction of Data Center Networks using Automatically Generated Simulation Models Piotr Rygielski, Samuel Kounev, Phuoc Tran-Gia Chair of Software Engineering University of Würzburg http://se.informatik.uni-wuerzburg.de/ SIMUtools2015, Athens, Greece, 25.08.2015
2 Piotr Rygielski Motivation What if Current performance known monitoring. Goal: predict performance after a change. (dst_ip>*.*.*.128)? port1 : port0; What if (src_tcp==80 && src_tcp==443)? port1 : port0; What if
3 Piotr Rygielski Research Gap Solution time Addressed Models Simulation models End-to-end performance analysis not detailed enough Existing network models too coarse or too fine grained Analytical models Level of details Other approaches focus only on selected technologies/protocols Flexibility in modeling is missing Time overhead Black-box models Detailed simulations Accuracy
4 Piotr Rygielski Approach Performance model(s) Model transformation(s) Descriptive model Real network Model extraction
5 Piotr Rygielski Approach Real network script Performance models Structure model Traffic model Configuration model to QN to OMNeT++ DNI meta model (modeling language) single model to QPN to formulas to ns3 other... Model-to-model transformations
6 Piotr Rygielski Models and Transformations Input Transformation Descriptive Model Predictive Model User Input (Model Extraction) DNI Model Routing format conversion DNI2QPN DNI2mDNI DNI2OMNeT++ Queueing Petri Net Model (DNI) minidni Model OMNeT++ Model mdni2qpn Queueing Petri Net Model (mdni)
minidni Meta-Model When not enough data to build full DNI instance Very coarse-granular modeling Network 0..* Route start end 1..* Node connects 2 Link 0..* destination 1 0..* TrafficSource 1 NodePerf 1 LinkPerf 1 Workload 7 Piotr Rygielski
8 Piotr Rygielski DNI Meta-Model (short) DNI Meta-Model Structure model Traffic model Configuration model Node NetworkInterface Link PerformanceDescriptions SoftwareComponent TrafficSource Flow Workload Route ProtocolStack NetworkProtocol Start Wait Loop Stop Transmit Sequence
9 Piotr Rygielski Transformation mdni-to-qpn QPN model of a network node, e.g., Switch, Server (mdni) Aspects: None, Generator, Receiver, Traversal input forward-traversing-traffic output dummy -trafficsource color-generation dummy generation-delay traffic-source node
10 Piotr Rygielski Transformation mdni-to-qpn QPN model of a network link (mdni) Delays from Interfaces and links integrated in queueing place transmission-delay node node transmission-delay link
11 Piotr Rygielski Transformation mdni-to-qpn
12 Piotr Rygielski Transformations - comparison
13 Piotr Rygielski Case study SBUS/PIRATES Traffic Management System Induction Loops GPS Sensors Traffic eras Traffic Light Sensors http://www.cl.cam.ac.uk/research/time/
14 Piotr Rygielski Case study SBUS/PIRATES
Case study SBUS/PIRATES...... Network...... 15 Piotr Rygielski.........
16 Piotr Rygielski Model Calibration simulated network simulation intergeneration time ON OFF generation delay = 0 ON t t network software ON OFF ON t generation delay simulation think time software think time t
17 Piotr Rygielski Experiment - Hardware VM 4.1 S4 S2 VM 5.1 S9 S8 S7 S5 S3 S6 VM 6.1 SW1 SW2 SW3 S1
Results Prediction Accuracy 18 Piotr Rygielski Motivation & Approach DNI & Transformations Current Focus Planning
Results Simulation Time Dumbbell topology Simulation duration [s] 500 400 300 200 100 0 100 90 OMNeT++ (30s) DNI QPN minidni QPN 80 70 60 50 40 30 20 10 Traffic intensity (think time) [ms] 19 Piotr Rygielski Motivation & Approach DNI & Transformations Current Focus Planning
Results Simulation Time Dumbbell topology Simulation duration [s] 900 800 700 600 500 400 300 200 100 0 OMNeT++ (30s) DNI QPN minidni QPN 10 20 30 40 50 60 70 80 90 100 Number of nodes 20 Piotr Rygielski Motivation & Approach DNI & Transformations Current Focus Planning
Conclusions Automatically generated three predictive models Prediction errors up to 18% for DNI (fully automatic process) minidni-qpn: accuracy loss (~4%) with speedup up to 300x Support for network virtualization in DNI (SDN planned) Model calibration is important. Modeling support tools needed 21 Piotr Rygielski Motivation & Approach DNI & Transformations Current Focus Planning
Thank You! Code & more info: http://go.uni-wuerzburg.de/aux piotr.rygielski@uni-wuerzburg.de http://se.informatik.uni-wuerzburg.de