A cheap, simple framework for distributed MATSim

• 300+ lines, 800+ routes, 2000+ stops, 70k+ links
• Needs PT on the network because PT a dominant mode here, detailed PT analysis essential
• Initially plans have negative scores because of low car availability, stranded agents
• Large solution space -> long runs; i.e. 2 days on 24 cores
• Long initialization time: 2hrs on our 24 core server

• QSim is limiting factor in performance.
• Increase QSim performance with increasing cores
• BUT You can only improve QSim by so much (tightly coupled, simple network decomposition)

• Reduce iterations (QSim expensive, limited parallellization)
• Reduce time per iteration, initialization (Large networks, transit routing)
• Increase agent plan memory
• Increase sample size (esp. important for transit: scaleability, error-prone because have to change many configurations)
• Use cheap computation, e.g. amazon EC2 or office network
• i.e. remove CPU, RAM limitations of single SMP

• Monolithic replanning modules are hard to maintain in an evolving project
• Need flexibility to grow with main MATSim project, easily integrate new functionality
• Produce comparable results to MATSim

• Transit PSim
• Plan and person serialization
• Load balancing

• Operation modes: serial vs parallel (parallel implies operating on information one iteration older)

• Full transit performance transmission

• Stochastic boarding model for PSim

• Works with most modules

• Works, in principle
• Parallel is preferable

• Increases the number of mutation combinations performed on a plan
• Co-evolutionary algorithm relies on mutation rate and number of iterations
• Number of unique mutation combinations performed on plans is multinomial distribution
• Each plan can be said to have a 'genome' of mutations, i.e. I000A001B003A004G010
• Standard MATSim can increase any given genome by a maximum of 1 letter

• How does genome length relate to exploration of solution space?

• Can genomes possibly tell us something about scenario characteristics/complexity?

• Can genome analysis inform effective replanning strategies, e.g by varying replanning rates or chaining strategies based on effective combinations ('genes') observed in the genome?

Design

• Make PSim a replanning module, no regeneration of data structures for replanning

Performance tests

• Number of slaves vs PSim iters vs number of threads per slave vs time

New strategies

• Diversification, large agent memories, randomized routing

Ensemble simulations

• Instead of a single master, have multiple masters running differrent combinations of agent plans
• Requires new modes of analysis