part2: software hardware architectureathena.ecs.csus.edu/~yaoz/pdf/abmpart2-softwarehardware.pdf ·...

Part2: Software & hardware architecture

1NYBPM Users Group Meeting, June 8, 2011

Topics Software architecture

Data structures Algorithms Distributed computing Outputs and visualization Outputs and visualization

Hardware architecture ARC ABM example ARC ABM example


CT‐RAMP: Software Design Concept

SANDAG•UEC spreadsheets•Market definitions

CT‐RAMP:•Model Flow•Model Components•Inputs/Outputs

Common Modeling Framework:•Matrix Classes

Inputs/Outputs

•LogitModel•Utility Expression Calculator

NYBPM Users Group Meeting, June 8, 2011 3

Common Modeling Framework A library of tools for building transport and land‐use modelsW i i h J i l Written in the Java programming language

Open source (Apache public license)C ll b i Collaborative

Currently used by over 30 clients


Wh J ?Why Java?

Java is a fully Object‐Oriented Programming (OOP) Language

Java is easy to learn and use Java encourages good software design Java natively supports multi‐threading Java is architecture‐neutral


OOP – Data Structures

Household 1

Synthetic PopulationOOP Data Structures

P 2P 1 P 3

AttributesHousehold SizeAutos OwnedNumber of WorkersHousehold Income

Person 2Person 1 Person 3Household IncomeHousehold LocationComposition

AttributesAgeGender

Tour 1 Tour 2 Tour 3 Tour 4 Tour 5Employment StatusStudent StatusWork LocationSchool Location

AttributesNumber ofNumber of Stops

Tour ModePersons

on TourPrimary

Choice Models


yDestination

Interaction between models and dataInteraction between models and data

HouseholdDataManager

Household 1

Person 2Person 1 Person 3

AttributesHousehold SizeAutos OwnedNumber of WorkersHousehold IncomeHousehold Location

Tour 1 Tour 2 Tour 3 Tour 4 Tour 5

Household LocationComposition

AttributesAgeGenderEmployment StatusStudent StatusWork LocationSchool Location

AttributesNumber ofNumber of Stops

Tour ModePersons

on TourPrimary

Destination


CMF Tools – Matrix Package Read/write to/from all major software (TransCAD, Cube, Emme, etc)

Matrix calculations Matrix calculations Random access (skims in memory, sparse matrices) N‐dimensional matrix, iterative proportional fittingd e s o a at , te at ve p opo t o a tt g


CMF Tools – Model Package Create and apply discrete choice models Flexible in specification of nesting structures “Interface” pattern used – any object can be an alternativeE i d b i f Extensive debugging features


CMF T l M d l P kCMF Tools – Model Package/** A simple mode choice model*/LogitModel.add()

public class MyModeChoiceModel {

public runModel(){

//instantiate modes

g ()takes a Mode

getUtility() solves logit model, returns logsum

//instantiate modesDriveAlone driveAlone = new DriveAlone();Transit transit = new Transit();

//instantiate model

g

//LogitModel model = new LogitModel;

//add modes to modelmodel.add(driveAlone);m d l dd(tr n it);

Uses Monte Carlo to select alternative according to logit

b bili i d model.add(transit);

//calculate utilitiesdouble logsum = model.getUtility();

probabilities and returns it.


//choose ModeMode chosenMode = (Mode) model.chooseAlternative();

}}

CMF Tools – Calculator Package Activity‐based models typically utilize many logit choice models, some with many alternatives

Traditional software relies on hard‐coded utility equationsTraditional software relies on hard coded utility equations Inefficient ‐ Programmer responsible for coding utility equations Inflexible – Requires programmer to change equations and

recompilerecompile Imperfect – Only one person typically reviews equations, which

increases probability of bugs Utility Expression Calculator (UEC) developed to overcome Utility Expression Calculator (UEC) developed to overcome these limitations


CMF Tools – Calculator Package

The UEC is a Java package that reads and interprets an Excel workbook containing a interprets an Excel workbook containing a logit model specification and its inputs

The UEC solves the utility equations for a y qgiven decision‐maker

The UEC “opens up” the model specification p p p– anyone can edit the spreadsheets, change inputs & parameters, check that the model is

l ifi d


properly specified, etc.

UEC ‐ Data PageTable Data: CSV files of zonal, household, or person data

Matrix Data: Trip tables or level-of-service skims in zone-f t (TPPLUS TRANSCAD EMME2 d/zone format (TPPLUS, TRANSCAD, EMME2, and/or

BINARY formats)

(Sparse matrices can be compressed in memory)


UEC ‐Model Page ‐ Tokens


UEC – Model Page ‐ Utilities


Threading

Quad‐core Intel Box with 4 GB RAM per process

Workplace Location Choice

Thread 1: households 1‐50,000

Household Data

Manager

Thread 2: households 50,001‐100,000Thread 3: households 100,001‐150,000


Thread 4: households 150,001‐200,000

Distribution

MasterNode

TaskScheduler Results

ResultsWriter


Grid Computing via the Java p gParallel Processing Framework

ABM Outputs

HID PID TID PUR MOD SB SA OTAZ DTAZ S1TAZ S2TAZ TLOR TLDS1 1 1 2 1 0 1 943 987 0 964 1 3

Household Data, Person Data, Tour/Trip List

1 1 1 2 1 0 1 943 987 0 964 1 3 1 1 2 1 2 1 0 943 731 856 0 3 3 1 2 1 4 1 0 0 943 952 0 0 1 2 1 3 1 2 4 1 1 943 565 698 982 1 2

Work Trip Frequency Distribution:Auto Ownership 1, Income Group 1-2

Estimated vs. Observed

4.0%

5.0%

6.0%

cy

Observed

Trip Tables

0.0%

1.0%

2.0%

3.0%

5 8 11 14 17 20 23 26 29 32 35 38 41 44 47 50 53 56 59

Peak Highway Travel Time (minutes)

Freq

uenc Estimated


Other SummariesMaps, Graphics Assignment

ABMVIZ: AB Model Visualization Dashboard

Developed in Adobe Flex/Flash Design goal: Intuitive + Flexible

F t Q i T bl d L t f Vi l Features Queries, Tables, and Lots of Visuals Bar Charts +

Four Types of Maps Time Use Tour Tracing Tree MapsTree Maps Radar Charts Bar Chart


ARC ABM Reporting Database

Stores ABM Output Tables, Travel Time Skim Data, and Select Inputs By Scenario Around 15 GBs per scenario Around 5–10 scenarios at a time in the DB Can load survey data into DB as a scenario for comparison purposes

Needs to be fast – not a traditional transactional DB, more a data warehouse Use de‐normalized data to avoid millions of joins each time for

dynamic visualization dashboardR d l i i l llb k f i d Read‐only; minimal rollback features required

Parallel query functionality (SQL Server) Needs excellent query caching functionalityU ARC E t i SQL S ith Li d P Use ARC Enterprise SQL Server with 4 Licensed Processors

ABMVIZ QueriesInteractive query builder with lots of default queriesDifference

alternatives

Q

Aggregations

Edit SQL if desired

Aggregations

Save Tables to Save Tables to Clipboard, Excel

Bar Chart and MapBar Chart and MapCan interactively hide/show bars and redraw plot Select field to visualize

Cycle thru fields anredraw map

Zoom in/out with “+”/”‐” or mouse wheel


Ti S t T li b I & P TTime Spent Traveling by Income & Person Type

NYBPM Users Group Meeting, June 8, 201124

T it Rid b A B F tTransit Riders by Age – Base vs. Future

NYBPM Users Group Meeting, June 8, 201125

Time UseCan change person type

Share of time use by purpose and person type


Tracing of Activities/Tours


Performance Measures with Radar Charts Compares independent measures across entities Compares independent measures across entities

Persons Not At Home By TAZ and Hour


Special Reports for Pricing Number of high a users (persons ehicles) Number of highway users (persons, vehicles):

Toll users: Captive

AM Peak Toll Trips by Time SavedCaptive

By choice

Non‐toll users: Captive By choice

Toll revenue Toll revenue Equity analysis


ARC HOV2HOT Conversion StudyARC HOV2HOT Conversion StudyToll Share of Toll‐Eligible Trips by Time Saved


ARC HOV2HOT Conversion StudyARC HOV2HOT Conversion StudyDiurnal Distribution of Toll versus Non‐Toll Work Tours


ARC ABM Hardware and Software Setup• Three Windows Server 2003

64bit Machines:D l Q d C I t l X • Dual Quad Core Intel Xeon X5570 2.93 GHz with Hyper‐Threading 16 threads

• 32 GB of RAM• Cube Voyager + 8 seat Cube

Cluster license

• Total cost ~ $30,000 in 2009

Implementation Design Goals • Overnight run time Model Relevance

• Around 16 hours• Requires distribution and threadingRequires distribution and threading• Model runtime is roughly proportional to population size• Network skimming and assignment procedures are still proportional to

the squared number of TAZs: ~50% or more of total model runtime

• Commodity hardware Minimize total lifetime cost• Hardware available today from common vendors; reasonably priced

• Easy to Setup and Use Staff acceptance• Not too complicated to setup, run, debug, etcp p g

ARC ABM System DesignARC ABM System Design


Distributing and Threading CT‐RAMPM i C b i t ll th JPPF li t t t t CT RAMP• Main Cube script calls the JPPF client to start CT‐RAMP

• ~1.76 million households split into 880 tasks of 2000 HHs• CT‐RAMP data managed through:g g

• Household Manager – manages all HH and person data into RAM for quick I/O

• Matrix Manager – reads all the matrix data into RAM for quick I/Og q• Run a sample of HHs to save time: 33% 50% 100%• HHs and Persons store a random number seed to avoid

random number sequence order of processing problemsrandom number sequence order of processing problems• SPEED UP = 9X

ARC ABM Run Times

3325

0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

Network Prep, Truck Model, Initial Skims

ARC ABM Run Times (min)

1400

24

25

112

6

II Demand with CT‐Ramp (33% Sample)

Convert Trip Lists to Demand Matrices

No Threading/Distribution (8 processors, 16GB RAM, 1 Computer)

Threaded and Distributed (24 processors, 48GB RAM, 3 computers)

170

2100

36

80

165

9

Highway & Transit Assignment & Skimming



170

4200

52

75

310



C T i Li D d M i 52

173

437

13

75

100



Highway Assignment (AM, PM, MD, NT)


8795970Total

part2: software hardware architectureathena.ecs.csus.edu/~yaoz/pdf/abmpart2-softwarehardware.pdf ·...

Documents