Fujimoto ICCS 6 15 Final
Efficient Execu+on of Replicated Transporta+on Simula+ons with Uncertain Vehicle Trajectories* Philip Pecher, Michael Hunter, and Richard Fujimoto Georgia Ins;tute of Technology * Supported by AFOSR Grant FA9550-‐13-‐1-‐0100
Outline
• Preliminaries and Problem Descrip;on • Simula;on Algorithm (superimposed simula;ons) • Experimental Evalua;on
Mo;va;ng Problem: Vehicle Tracking • Track a target vehicle in a complex urban environment • Do not have con;nuous monitoring; repeatedly acquire and reacquire target vehicle • Fixed-‐posi;on sensors (cameras) augmented with mobile sensors, crowd-‐sourced informa;on • Assume target is not evasive
DDDAS Solu;on Approach Adapt: reposi;on sensors or target data queries in light of predicted loca;ons ?
?
Sense: obtain loca;on of target vehicle from fixed or mobile sensors, crowd-‐sourced data, etc.
?
Predict: construct a probability map indica;ng the likelihood the target vehicle will be at different loca;ons at some ;me in the future
Predic;on Using Replicated Simula;ons Approach • Analy;cs for des;na;on predic;on [Pecher et al., 2014] • Use replicated simula;ons to predict possible loca;ons at ;me T+∆T based on alternate routes the vehicle might take (or other varia;ons of target vehicle behavior) – K replica;ons; iden;cal except for target vehicle behavior
Goals • Accelerate execu;on of replica;on simula;ons • Produce exactly the same results as a brute-‐force replicated simula;on experiment
Observa;ons • Much of the computa;on among the different replica;ons will be the same (only the target vehicle varies); these computa;ons need only be performed once • Only the state of the target vehicle is of interest; errors in the simula;on that do not impact the state of the target vehicle can be tolerated!
Related Work • Standard Clock for replicated simula;ons [Vakili 1992] – Exploits shared computa;ons among replica;ons
• Staged Simula;ons [Walsh, Sirer 2003]
– More general computa;on sharing among simula;ons
• Cloning parallel and distributed simula;ons [Hybinede, Fujimoto 2001; Chen, Turner, Cai, Gan, Low 2005; Chen 2012] – Complementary; improve efficiency of crea;ng replica;ons
• Updateable Simula;ons [Ferenci, et al. 2002]
– Requires genera;ng a log of the simula;on prior to execu;on
Overview of Algorithm • Preliminaries – One target vehicle per replica;on; called a Virtual Vehicle (VV) – Other vehicles called Model Vehicles (MVs)
• Create one superimposed execu2on including – One execu;on of MVs; these behave oblivious to the VVs – K VVs; each VV is oblivious to other VVs, but interacts with MV in accordance with the model rules (e.g., avoid collisions)
• The behavior of some MVs are wrong (because they ignore the VVs); tag these MVs as hazards – Hazards do not necessarily compromise simula;on results because we are only interested in the posi;on of VVs
• If a hazard does affect a VV, the posi;on of the VV is incorrect; resort to rollback and cloning a new replica;on
Nagel-‐Schreckenberg Transporta;on Model Cellular Automata 1. The speed of the vehicle is incremented if the vehicle is not at the maximum speed. 2. The current vehicle speed is compared to the number of empty cells in front of the vehicle. If the number of empty cells is smaller, the speed is reduced to the empty cell count (in order to avoid a poten;al collision). 3. If the vehicle's speed is posi;ve, it is decremented with probability p (this is a predefined parameter). 4. The vehicle advances forward by the number of cells equal to the corresponding speed.
Superimposed Simula;on Overlay replicated simula;ons so common computa;ons are only performed once MV movement is oblivious to (ignores) virtual vehicles
Hazards and Tags
Hazard (tag=1)
• A hazard is a MV whose state is incorrect in one or more replica;ons • Hazards do not compromise the simula;on, so long as they do not impact the state of the target vehicle (virtual vehicle) • Each hazard is tagged to indicate incorrect replica;ons
Algorithm Rules • In the superimposed simula;on
– MVs interact with other MVs but ignore VVs, except for hazard detec;on – VVs interact with MVs, as usual
• Hazard crea;on: If a MV MV1 interacts with a VV VVk, then MV1 is tagged as a “hazard” for replica;on k (indicates the superimposed simula;on is in error for replica;on k) • Hazard spread: If a second vehicle MV2 interacts with MV1, MV2 is also tagged as a hazard in replica;on k • Rollback: If a VV interacts with a tagged MV, we must roll back, replicate (clone) the simula;on
Outline
• Preliminaries and Problem Descrip;on • Simula;on Algorithm (superimposed simula;ons) • Experimental Evalua;on
Experimental Evalua;on • Compare brute-‐force replica;on approach with superimposed simula;ons for NS traffic model • Two road networks – Manhadan-‐style grid with long bi-‐direc;onal road segments (450 m) – Road network based on a sec;on in Atlanta Georgia (US) (mean road segment length: 135 m)
• Each simula;on runs for 500 ;mesteps.
Vehicle Movement • Vehicles take the shortest path to randomly selected intersec;ons. • Vehicles are randomly generated at the outermost points of the map boundary. – “Low traffic”: 0.1 spawn probability per source per ;ck – “High traffic”: 0.5 spawn probability per source per ;ck
• Default values for Model Vehicles: – N-‐S decelera;on parameter p = 0.5 – Maximum speed 3 cells/;mestep
• Varying parameters for Virtual Vehicles (up to 64 configura;ons):
– Decelera+on Prob. = – Max. Speed = cells/;mestep – Des+na+on =
Manhadan Grid
• Intersec;on-‐to-‐intersec;on distance = 100 cells • Bidirec;onal 2-‐lane roads • Vehicles spawn at the red circles in the above illustra;on.
Execu;on Time: Manhadan Network 120
100
Speedup=1.65
CPU Time (s)
80
0.1 VVEM 60
0.1 BFEM 0.5 VVEM 0.5 BFEM
40
Speedup=5.8
20
0 1
2
4
8
16
32
64
Number of Replica+ons
• Brute force replica;ons vs. superimposed simula;ons • Speedup increases with more replica;ons • Beder speedup in networks with light traffic loads
Speedup vs. Rela;ve Speed Differences 18 16 14 12 10
High Decel, Low Speed Low Decel, Low Speed
8
High Decel, High Speed Low Decel, High Speed
6
(VV speed)
4 2 0 High Traffic, V=5
High Traffic, V=3
Low Traffic, V=5
Low Traffic, V=3
• Beder speedup with lower maximum vehicle speeds • Beder speedup when VVs are “fast” rela;ve to MVs
Atlanta Traffic Scenario
Execu;on Time: Atlanta Network 18 16 14
CPU Time (s)
12 10
0.1 VVEM 0.1 BFEM
8
0.5 VVEM 0.5 BFEM
6 4 2 0 1
2
4
8
16
32
64
Number of Replica+ons
Results comparable to synthe;c “Manhadan” network
Atlanta: Speedup vs. Rela;ve Speed Differences
14
12
10
High Decel, Low Speed
8
Low Decel, Low Speed High Decel, High Speed
6
Low Decel, High Speed 4
2
0 High Traffic, V=5
High Traffic, V=3
Low Traffic, V=5
Low Traffic, V=3
Conclusions and Open Ques;ons Conclusions • Superimposed simula;ons provides a means to accelerate replicated execu;ons of similar NS traffic simula;ons • Up to order-‐of-‐magnitude speedup observed • Performance depends significantly on specifics of the model (traffic conges;on, speed of target rela;ve to other vehicles, etc.) Open Ques;ons • Approximate techniques • Generaliza;on to other types of traffic models • Generaliza;on to other applica;ons • Incorpora;on of incremental replica;on (cloning) concepts • Parallel implementa;on • Extension to tracking mul;ple vehicles
Outline
• Preliminaries and Problem Descrip;on • Simula;on Algorithm (superimposed simula;ons) • Experimental Evalua;on
Mo;va;ng Problem: Vehicle Tracking • Track a target vehicle in a complex urban environment • Do not have con;nuous monitoring; repeatedly acquire and reacquire target vehicle • Fixed-‐posi;on sensors (cameras) augmented with mobile sensors, crowd-‐sourced informa;on • Assume target is not evasive
DDDAS Solu;on Approach Adapt: reposi;on sensors or target data queries in light of predicted loca;ons ?
?
Sense: obtain loca;on of target vehicle from fixed or mobile sensors, crowd-‐sourced data, etc.
?
Predict: construct a probability map indica;ng the likelihood the target vehicle will be at different loca;ons at some ;me in the future
Predic;on Using Replicated Simula;ons Approach • Analy;cs for des;na;on predic;on [Pecher et al., 2014] • Use replicated simula;ons to predict possible loca;ons at ;me T+∆T based on alternate routes the vehicle might take (or other varia;ons of target vehicle behavior) – K replica;ons; iden;cal except for target vehicle behavior
Goals • Accelerate execu;on of replica;on simula;ons • Produce exactly the same results as a brute-‐force replicated simula;on experiment
Observa;ons • Much of the computa;on among the different replica;ons will be the same (only the target vehicle varies); these computa;ons need only be performed once • Only the state of the target vehicle is of interest; errors in the simula;on that do not impact the state of the target vehicle can be tolerated!
Related Work • Standard Clock for replicated simula;ons [Vakili 1992] – Exploits shared computa;ons among replica;ons
• Staged Simula;ons [Walsh, Sirer 2003]
– More general computa;on sharing among simula;ons
• Cloning parallel and distributed simula;ons [Hybinede, Fujimoto 2001; Chen, Turner, Cai, Gan, Low 2005; Chen 2012] – Complementary; improve efficiency of crea;ng replica;ons
• Updateable Simula;ons [Ferenci, et al. 2002]
– Requires genera;ng a log of the simula;on prior to execu;on
Overview of Algorithm • Preliminaries – One target vehicle per replica;on; called a Virtual Vehicle (VV) – Other vehicles called Model Vehicles (MVs)
• Create one superimposed execu2on including – One execu;on of MVs; these behave oblivious to the VVs – K VVs; each VV is oblivious to other VVs, but interacts with MV in accordance with the model rules (e.g., avoid collisions)
• The behavior of some MVs are wrong (because they ignore the VVs); tag these MVs as hazards – Hazards do not necessarily compromise simula;on results because we are only interested in the posi;on of VVs
• If a hazard does affect a VV, the posi;on of the VV is incorrect; resort to rollback and cloning a new replica;on
Nagel-‐Schreckenberg Transporta;on Model Cellular Automata 1. The speed of the vehicle is incremented if the vehicle is not at the maximum speed. 2. The current vehicle speed is compared to the number of empty cells in front of the vehicle. If the number of empty cells is smaller, the speed is reduced to the empty cell count (in order to avoid a poten;al collision). 3. If the vehicle's speed is posi;ve, it is decremented with probability p (this is a predefined parameter). 4. The vehicle advances forward by the number of cells equal to the corresponding speed.
Superimposed Simula;on Overlay replicated simula;ons so common computa;ons are only performed once MV movement is oblivious to (ignores) virtual vehicles
Hazards and Tags
Hazard (tag=1)
• A hazard is a MV whose state is incorrect in one or more replica;ons • Hazards do not compromise the simula;on, so long as they do not impact the state of the target vehicle (virtual vehicle) • Each hazard is tagged to indicate incorrect replica;ons
Algorithm Rules • In the superimposed simula;on
– MVs interact with other MVs but ignore VVs, except for hazard detec;on – VVs interact with MVs, as usual
• Hazard crea;on: If a MV MV1 interacts with a VV VVk, then MV1 is tagged as a “hazard” for replica;on k (indicates the superimposed simula;on is in error for replica;on k) • Hazard spread: If a second vehicle MV2 interacts with MV1, MV2 is also tagged as a hazard in replica;on k • Rollback: If a VV interacts with a tagged MV, we must roll back, replicate (clone) the simula;on
Outline
• Preliminaries and Problem Descrip;on • Simula;on Algorithm (superimposed simula;ons) • Experimental Evalua;on
Experimental Evalua;on • Compare brute-‐force replica;on approach with superimposed simula;ons for NS traffic model • Two road networks – Manhadan-‐style grid with long bi-‐direc;onal road segments (450 m) – Road network based on a sec;on in Atlanta Georgia (US) (mean road segment length: 135 m)
• Each simula;on runs for 500 ;mesteps.
Vehicle Movement • Vehicles take the shortest path to randomly selected intersec;ons. • Vehicles are randomly generated at the outermost points of the map boundary. – “Low traffic”: 0.1 spawn probability per source per ;ck – “High traffic”: 0.5 spawn probability per source per ;ck
• Default values for Model Vehicles: – N-‐S decelera;on parameter p = 0.5 – Maximum speed 3 cells/;mestep
• Varying parameters for Virtual Vehicles (up to 64 configura;ons):
– Decelera+on Prob. = – Max. Speed = cells/;mestep – Des+na+on =
Manhadan Grid
• Intersec;on-‐to-‐intersec;on distance = 100 cells • Bidirec;onal 2-‐lane roads • Vehicles spawn at the red circles in the above illustra;on.
Execu;on Time: Manhadan Network 120
100
Speedup=1.65
CPU Time (s)
80
0.1 VVEM 60
0.1 BFEM 0.5 VVEM 0.5 BFEM
40
Speedup=5.8
20
0 1
2
4
8
16
32
64
Number of Replica+ons
• Brute force replica;ons vs. superimposed simula;ons • Speedup increases with more replica;ons • Beder speedup in networks with light traffic loads
Speedup vs. Rela;ve Speed Differences 18 16 14 12 10
High Decel, Low Speed Low Decel, Low Speed
8
High Decel, High Speed Low Decel, High Speed
6
(VV speed)
4 2 0 High Traffic, V=5
High Traffic, V=3
Low Traffic, V=5
Low Traffic, V=3
• Beder speedup with lower maximum vehicle speeds • Beder speedup when VVs are “fast” rela;ve to MVs
Atlanta Traffic Scenario
Execu;on Time: Atlanta Network 18 16 14
CPU Time (s)
12 10
0.1 VVEM 0.1 BFEM
8
0.5 VVEM 0.5 BFEM
6 4 2 0 1
2
4
8
16
32
64
Number of Replica+ons
Results comparable to synthe;c “Manhadan” network
Atlanta: Speedup vs. Rela;ve Speed Differences
14
12
10
High Decel, Low Speed
8
Low Decel, Low Speed High Decel, High Speed
6
Low Decel, High Speed 4
2
0 High Traffic, V=5
High Traffic, V=3
Low Traffic, V=5
Low Traffic, V=3
Conclusions and Open Ques;ons Conclusions • Superimposed simula;ons provides a means to accelerate replicated execu;ons of similar NS traffic simula;ons • Up to order-‐of-‐magnitude speedup observed • Performance depends significantly on specifics of the model (traffic conges;on, speed of target rela;ve to other vehicles, etc.) Open Ques;ons • Approximate techniques • Generaliza;on to other types of traffic models • Generaliza;on to other applica;ons • Incorpora;on of incremental replica;on (cloning) concepts • Parallel implementa;on • Extension to tracking mul;ple vehicles
