tutorial Parallel Distributed Simulation Systems From ChandyMisra to

tutorial: Parallel & Distributed Simulation Systems: From Chandy/Misra to the High Level Architecture and Beyond Richard M. Fujimoto College of Computing Georgia Institute of Technology Atlanta, GA 30332 -0280 fujimoto@cc. gatech. edu

References R. Fujimoto, Parallel and Distributed Simulation Systems, Wiley Interscience, 2000. (see also http: //www. cc. gatech. edu/classes/AY 2000/cs 4230_spring) HLA: F. Kuhl, R. Weatherly, J. Dahmann, Creating Computer Simulation Systems: An Introduction to the High Level Architecture for Simulation, Prentice Hall, 1999. (http: //hla. dmso. mil)

Outline Part I: Introduction Part II: Time Management Part III: Distributed Virtual Environments

Parallel and Distributed Simulation Parallel simulation involves the execution of a single simulation program on a collection of tightly coupled processors (e. g. , a shared memory multiprocessor). simulation model P P M P M parallel processor Replicated trials involves the execution of several, independent simulations concurrently on different processors Distributed simulation involves the execution of a single simulation program on a collection of loosely coupled processors (e. g. , PCs interconnected by a LAN or WAN).

Reasons to Use Parallel / Distributed Simulation Enable the execution of time consuming simulations that could not otherwise be performed (e. g. , simulation of the Internet) • Reduce model execution time (proportional to # processors) • Ability to run larger models (more memory) Enable simulation to be used as a forecasting tool in time critical decision making processes (e. g. , air traffic control) • Initialize simulation to current system state • Faster than real time execution for what-if experimentation • Simulation results may be needed in seconds Create distributed virtual environments, possibly including users at distant geographical locations (e. g. , training, entertainment) • Real-time execution capability • Scalable performance for many users & simulated entities

Geographically Distributed Users/Resources Geographically distributed users and/or resources are sometime needed • Interactive games over the Internet • Specialized hardware or databases Server architecture Cluster of workstations on LAN Distributed architecture Distributed computers WAN interconnect LAN interconnect

Stand-Alone vs. Federated Simulation Systems Stand-Alone Simulation System Federated Simulation Systems Simulator 1 Process 1 Simulator 2 Simulator 3 Process 2 Process 3 Process 4 Parallel simulation environment Homogeneous programming environment • simulation language Run Time Infrastructure-RTI (simulation backplane) • federation set up & tear down • synchronization, message ordering • data distribution Interconnect autonomous, heterogeneous simulators • interface to RTI software

Principal Application Domains Parallel Discrete Event Simulation (PDES) • Discrete event simulation to analyze systems • Fast model execution (asfast-as-possible) • Produce same results as a sequential execution • Typical applications – – Telecommunication networks Computer systems Transportation systems Military strategy and tactics Distributed Virtual Environments (DVEs) • Networked interactive, immersive environments • Scalable, real-time performance • Create virtual worlds that appear realistic • Typical applications – Training – Entertainment – Social interaction

Historical Perspective High Performance Computing Community Chandy/Misra/Bryant algorithm Time Warp algorithm second generation algorithms making it fast and easy to use early experimental data 1975 1980 1985 1990 SIMulator NETworking (SIMNET) (1983 -1990) Defense Community 2000 High Level Architecture (1996 - today) Distributed Interactive Simulation (DIS) Aggregate Level Simulation Protocol (ALSP) (1990 - 1997 ish) Dungeons and Dragons Board Games Multi-User Dungeon (MUD) Adventure Games (Xerox PARC) Internet & Gaming Community 1995 Multi-User Video Games

Part II: Time Management • • Parallel discrete event simulation Conservative synchronization Optimistic synchronization Time Management in the High Level Architecture

Time • physical system: the actual or imagined system being modeled • simulation: a system that emulates the behavior of a physical system main() {. . . double clock; . . . } physical system simulation • physical time: time in the physical system – Noon, December 31, 1999 to noon January 1, 2000 • simulation time: representation of physical time within the simulation – floating point values in interval [0. 0, 24. 0] • wallclock time: time during the execution of the simulation, usually output from a hardware clock (e. g. , GPS) – 9: 00 to 9: 15 AM on September 10, 1999

Paced vs. Unpaced Execution Modes of execution • As-fast-as-possible execution (unpaced): no fixed relationship necessarily exists between advances in simulation time and advances in wallclock time • Real-time execution (paced): each advance in simulation time is paced to occur in synchrony with an equivalent advance in wallclock time • Scaled real-time execution (paced): each advance in simulation time is paced to occur in synchrony with S * an equivalent advance in wallclock time (e. g. , 2 x wallclock time) Here, focus on as-fast-as-possible; execution can be paced to run in real-time (or scaled real-time) by inserting delays

Discrete Event Simulation Fundamentals Discrete event simulation: computer model for a system where changes in the state of the system occur at discrete points in simulation time. Fundamental concepts: • system state (state variables) • state transitions (events) • simulation time: totally ordered set of values representing time in the system being modeled (physical system) • simulator maintains a simulation time clock A discrete event simulation computation can be viewed as a sequence of event computations Each event computation contains a (simulation time) time stamp indicating when that event occurs in the physical system. Each event computation may: (1) modify state variables, and/or (2) schedule new events into the simulated future.

A Simple DES Example H 1 H 0 Time Event 0 H 0 Send Pkt 100 Done • • Time Event 1 H 1 Recv Pkt 6 H 0 Retrans 100 Done Time Event 1 H 1 Send Ack 6 H 0 Retrans 100 Done Time Event 2 H 0 Recv Ack 6 H 0 Retrans 100 Done Time Event 2 H 0 Send Pkt 100 Done Simulator maintains event list Events processed in simulation time order Processing events may generate new events Complete when event list is empty (or some other termination condition)

Parallel Discrete Event Simulation A parallel discrete event simulation program can be viewed as a collection of sequential discrete event simulation programs executing on different processors that communicate by sending time stamped messages to each other “Sending a message” is synonymous with “scheduling an event”

Parallel Discrete Event Simulation Example Physical system physical process logical process interactions among physical processes time stamped event (message) LP (Subnet 1) Simulation Recv Pkt @15 LP (Subnet 2) all interactions between LPs must be via messages (no shared state) LP (Subnet 3)

The “Rub” Golden rule for each logical process: “Thou shalt process incoming messages in time stamp order!!!” (local causality constraint)

Parallel Discrete Event Simulation Example Physical system physical process logical process Simulation interactions among physical processes time stamped event (message) Safe to process? ? ? LP (Subnet 1) Recv Pkt @15 LP (Subnet 2) all interactions between LPs must be via messages (no shared state) LP (Subnet 3)

The Synchronization Problem Local causality constraint: Events within each logical process must be processed in time stamp order Observation: Adherence to the local causality constraint is sufficient to ensure that the parallel simulation will produce exactly the same results as the corresponding sequential simulation* Synchronization (Time Management) Algorithms • Conservative synchronization: avoid violating the local causality constraint (wait until it’s safe) – 1 st generation: null messages (Chandy/Misra/Bryant) – 2 nd generation: time stamp of next event • Optimistic synchronization: allow violations of local causality to occur, but detect them at runtime and recover using a rollback mechanism – Time Warp (Jefferson) – approaches limiting amount of optimistic execution * provided events with the same time stamp are processed in the same order as in the sequential execution

Part II: Time Management • • Parallel discrete event simulation Conservative synchronization Optimistic synchronization Time Management in the High Level Architecture

Chandy/Misra/Bryant “Null Message” Algorithm Assumptions • logical processes (LPs) exchanging time stamped events (messages) • static network topology, no dynamic creation of LPs • messages sent on each link are sent in time stamp order • network provides reliable delivery, preserves order Observation: The above assumptions imply the time stamp of the last message received on a link is a lower bound on the time stamp (LBTS) of subsequent messages received on that link H 1 H 2 9 8 2 H 3 5 4 H 3 logical process Goal: Ensure LP processes events in time stamp order one FIFO queue per incoming link

A Simple Conservative Algorithm A (executed by each LP): Goal: Ensure events are processed in time stamp order: WHILE (simulation is not over) wait until each FIFO contains at least one message remove smallest time stamped event from its FIFO process that event END-LOOP H 1 9 8 2 H 2 5 4 H 3 logical process • process time stamp 2 event • process time stamp 4 event • process time stamp 5 event • wait until a message is received from H 2 Observation: Algorithm A is prone to deadlock!

Deadlock Example H 1 (waiting on H 2) 15 10 H 2 (waiting on H 3) 7 H 3 (waiting on H 1) 9 8 A cycle of LPs forms where each is waiting on the next LP in the cycle. No LP can advance; the simulation is deadlocked.

Deadlock Avoidance Using Null Messages Break deadlock by having each LP send “null” messages indicating a lower bound on the time stamp of future messages it could send. 11 15 10 H 2 (waiting on H 3) H 1 (waiting on H 2) 8 7 H 3 (waiting on H 1) 9 8 Assume minimum delay between hosts is 3 units of simulation time • H 3 initially at time 5 • H 3 sends null message to H 2 with time stamp 8 • H 2 sends null message to H 1 with time stamp 11 • H 1 may now process message with time stamp 7

Deadlock Avoidance Using Null Messages Null Message Algorithm (executed by each LP): Goal: Ensure events are processed in time stamp order and avoid deadlock WHILE (simulation is not over) wait until each FIFO contains at least one message remove smallest time stamped event from its FIFO process that event send null messages to neighboring LPs with time stamp indicating a lower bound on future messages sent to that LP (current time plus lookahead) END-LOOP The null message algorithm relies on a “lookahead” (minimum delay).

Lookahead Creep 6. 0 H 1 (waiting on H 2) 7 7. 5 6. 5 15 10 H 2 (waiting on H 3) 5. 5 7. 0 H 3 (waiting on H 1) 9 8 0. 5 Assume minimum delay between hosts is 3 units of time • H 3 initially at time 5 • H 3 sends null message, time stamp 5. 5; H 2 sends null message, time stamp 6. 0 • H 1 sends null message, time stamp 6. 5; H 3 send null message, time stamp 7. 0 • H 2 sends null message, time stamp 7. 5; H 1 can process time stamp 7 message Five null messages to process a single event If lookahead is small, there may be many null messages!

Preventing Lookahead Creep: Next Event Time Information H 1 (waiting on H 2) 15 10 H 2 (waiting on H 3) 7 H 3 (waiting on H 1) 9 8 Observation: If all LPs are blocked, they can immediately advance to the time of the minimum time stamp event in the system

Lower Bound on Time Stamp No null messages, assume any LP can send messages to any other LP When a LP blocks, compute lower bound on time stamp (LBTS) of messages it may later receive; those with time stamp < LBTS safe to process H 3@6 (not blocked) 8 H 3 9 H 2@6. 1 (blocked) 10 H 2 15 7 transient H 1 ? (6, 10, 7) (assume zero lookahead) H 1@5 LBTS = min 5 6 7 8 9 10 11 12 13 Simulation time 14 15 16 Given a snapshot of the computation, LBTS is the minimum among • Time stamp of any transient messages (sent, but not yet received) • Unblocked LPs: Current simulation time + lookahead • Blocked LPs: Time of next event + lookahead

Lower Bound on Time Stamp (LBTS) LBTS can be computed asynchonously using a distributed snapshot algorithm (Mattern) LP 4 LP 3 LP 2 LP 1 cut message Past Future wallclock time cut point: an instant dividing process’s computation into past and future cut: set of cut points, one per process cut message: a message that was sent in the past, and received in the future consistent cut: cut + all cut messages cut value: minimum among (1) local minimum at each cut point and (2) time stamp of cut messages; non-cut messages can be ignored It can be shown LBTS = cut value

A Simple LBTS Algorithm LP 4 LP 3 LP 2 LP 1 cut message Past Future wallclock time Initiator broadcasts start LBTS computation message to all LPs Each LP sets cut point, reports local minimum back to initiator Account for transient (cut) messages • Identify transient messages, include time stamp in minimum computation – Color each LP (color changes with each cut point); message color = color of sender – An incoming message is transient if message color equals previous color of receiver – Report time stamp of transient to initiator when one is received • Detecting when all transients have been received – For each color, LPi keeps counter of messages sent (Sendi) and received (Receivei) – At cut point, send counters to initiator: # transients = (Sendi – Receivei) – Initiator detects termination (all transients received), broadcasts global minimum

Another LBTS Algorithm LP 4 LP 3 LP 2 LP 1 cut message Past Future wallclock time An LP initiates an LBTS computation when it blocks Initiator: broadcasts start LBTS message to all LPs LPi places cut point, report local minimum and (Sendi – Receivei) back to initiator Initiator: After all reports received if ( (Sendi – Receivei) = 0) LBTS = global minimum, broadcast LBTS value Else Repeat broadcast/reply, but do not establish a new cut

Synchronous Algorithms Barrier Synchronization: when a process reaches a barrier synchronization, it must block until all other processors have also reached the barrier. process 1 wallclock time process 2 process 3 process 4 - barrier wait - barrier - wait Synchronous algorithm DO WHILE (unprocessed events remain) barrier synchronization; flush all messages from the network LBTS = min (Ni + LAi); Ni = time of next event in LPi; LAi = lookahead of LPi all i S = set of events with time stamp ≤ LBTS process events in S end. DO Variations proposed by Lubachevsky, Ayani, Chandy/Sherman, Nicol

Topology Information ORD 4: 00 6: 00 LAX 0: 30 SAN 2: 00 JFK 10: 45 10: 00 Global LBTS algorithm is overly conservative: does not exploit topology information • Lookahead = minimum flight time to another airport • Can the two events be processed concurrently? – Yes because the event @ 10: 00 cannot affect the event @ 10: 45 • Simple global LBTS algorithm: – LBTS = 10: 30 (10: 00 + 0: 30) – Cannot process event @ 10: 45 until next LBTS computation

Distance Between LPs • Associate a lookahead with each link: LAB is the lookahead on the link from LPA to LPB – Any message sent on the link from LPA to LPB must have a time stamp of TA + LAB where TA is the current simulation time of LPA • A path from LPA to LPZ is defined as a sequence of LPs: LPA, LPB, …, LPY, LPZ • The lookahead of a path is the sum of the lookaheads of the links along the path • DAB, the minimum distance from LPA to LPB is the minimum lookahead over all paths from LPA to LPB • The distance from LPA to LPB is the minimum amount of simulated time that must elapse for an event in LPA to affect LPB

Distance Between Processes The distance from LPA to LPB is the minimum amount of simulated time that must elapse for an event in LPA to affect LPB 11 3 LPA LPB 4 1 3 1 4 2 LPC LPD 2 Distance Matrix: D [i, j] = minimum distance from LPi to LPj LPA LPB LPC LPD LPA 4 3 1 3 LPB 4 5 3 1 min (1+2, 3+1) LPC 3 6 4 2 LPD 5 4 2 4 13 15 • An event in LPY with time stamp TY depends on an event in LPX with time stamp TX if TX + D[X, Y] < TY • Above, the time stamp 15 event depends on the time stamp 11 event, the time stamp 13 event does not.

Computing LBTS Assuming all LPs are blocked and there are no transient messages: LBTSi=min(Nj+Dji) (all j) where Ni = time of next event in LPi 11 3 LPA LPB 4 1 3 1 4 2 LPC LPD 2 13 15 Distance Matrix: D [i, j] = minimum distance from LPi to LPj LPA LPB LPC LPD LPA 4 3 1 3 LPB 4 5 3 1 min (1+2, 3+1) LPC 3 6 4 2 LPD 5 4 2 4 LBTSA = 15 [min (11+4, 13+5)] LBTSB = 14 [min (11+3, 13+4)] LBTSC = 12 [min (11+1, 13+2)] LBTSD = 14 [min (11+3, 13+4)] Need to know time of next event of every other LP Distance matrix must be recomputed if lookahead changes

Example ORD 4: 00 6: 00 LAX 0: 30 SAN 2: 00 JFK 10: 45 10: 00 Using distance information: • DSAN, JFK = 6: 30 • LBTSJFK = 16: 30 (10: 00 + 6: 30) • Event @ 10: 45 can be processed this iteration • Concurrent processing of events at times 10: 00 and 10: 45

Lookahead problem: limited concurrency each LP must process events in time stamp order event without lookahead possible message OK to process LP D LP C with lookahead possible message LP B OK to process LP A not OK to process yet TA TA+LA Logical Time Each LP A using declares a lookahead value LA; the time stamp of any event generated by the LP must be ≥ TA+ LA • Used in virtually all conservative synchronization protocols • Relies on model properties (e. g. , minimum transmission delay) Lookahead is necessary to allow concurrent processing of events with different time stamps (unless optimistic event processing is used)

Speedup of Central Server Queueing Model Simulation Deadlock Detection and Recovery Algorithm (5 processors) Exploiting lookahead is essential to obtain good performance

Summary: Conservative Synchronization • Each LP must process events in time stamp order • Must compute lower bound on time stamp (LBTS) of future messages an LP may receive to determine which events are safe to process • 1 st generation algorithms: LBTS computation based on current simulation time of LPs and lookahead – Null messages – Prone to lookahead creep • 2 nd generation algorithms: also consider time of next event to avoid lookahead creep • Other information, e. g. , LP topology, can be exploited • Lookahead is crucial to achieving concurrent processing of events, good performance

Conservative Algorithms Pro: • Good performance reported for many applications containing good lookahead (queueing networks, communication networks, wargaming) • Relatively easy to implement • Well suited for “federating” autonomous simulations, provided there is good lookahead Con: • Cannot fully exploit available parallelism in the simulation because they must protect against a “worst case scenario” • Lookahead is essential to achieve good performance • Writing simulation programs to have good lookahead can be very difficult or impossible, and can lead to code that is difficult to maintain

Part II: Time Management • • Parallel discrete event simulation Conservative synchronization Optimistic synchronization Time Management in the High Level Architecture

Time Warp Algorithm (Jefferson) Assumptions • logical processes (LPs) exchanging time stamped events (messages) • dynamic network topology, dynamic creation of LPs OK • messages sent on each link need not be sent in time stamp order • network provides reliable delivery, but need not preserve order Basic idea: • process events w/o worrying about messages that will arrive later • detect out of order execution, recover using rollback H 1 9 8 2 5 4 H 2 H 3 logical process all available events (2, 4, 5, 8, 9) in time stamp order

Time Warp (Jefferson) Each LP: process events in time stamp order, like a sequential simulator, except: (1) do NOT discard processed events and (2) add a rollback mechanism Input Queue (event list) processed event 12 21 35 41 snapshot of LP state State Queue Output Queue (anti-messages) unprocessed event anti-message 12 12 19 42 18 straggler message arrives in the past, causing rollback Adding rollback: • a message arriving in the LP’s past initiates rollback • to roll back an event computation we must undo: – changes to state variables performed by the event; solution: checkpoint state or use incremental state saving (state queue) – message sends solution: anti-messages and message annihilation (output queue)

Anti-Messages 42 positive message anti-message 42 • Used to cancel a previously sent message • Each positive message sent by an LP has a corresponding antimessage • Anti-message is identical to positive message, except for a sign bit • When an anti-message and its matching positive message meet in the same queue, the two annihilate each other (analogous to matter and anti-matter) • To undo the effects of a previously sent (positive) message, the LP need only send the corresponding anti-message • Message send: in addition to sending the message, leave a copy of the corresponding anti-message in a data structure in the sending LP called the output queue.

Rollback: Receiving a Straggler Message 2. roll back events at times 21 and 35 2(a) restore state of LP to that prior to processing time stamp 21 event Input Queue (event list) 12 21 35 41 processed event unprocessed event State Queue snapshot of LP state anti-message Output Queue (anti-messages) 12 12 19 42 2(b) send anti-message 18 BEFORE 1. straggler message arrives in the past, causing rollback Input Queue (event list) 12 18 State Queue Output Queue (anti-messages) AFTER 21 35 41 5. resume execution by processing event at time 18 12 12 19

Processing Incoming Anti-Messages Case I: corresponding message has not yet been processed – annihilate message/anti-message pair Case II: corresponding message has already been processed – roll back to time prior to processing message (secondary rollback) – annihilate message/anti-message pair 1. anti-message arrives 3. Annihilate message and anti-message 27 42 42 2. roll back events (time stamp 42 and 45) 2(a) restore state processed event 45 55 unprocessed event snapshot of LP state anti-message 33 57 2(b) send anti-message may cause “cascaded” rollbacks; recursively applying eliminates all effects of error Case III: corresponding message has not yet been received – queue anti-message – annihilate message/anti-message pair when message is received

Global Virtual Time and Fossil Collection A mechanism is needed to: • reclaim memory resources (e. g. , old state and events) • perform irrevocable operations (e. g. , I/O) Observation: A lower bound on the time stamp of any rollback that can occur in the future is needed. Global Virtual Time (GVT) is defined as the minimum time stamp of any unprocessed (or partially processed) message or anti-message in the system. GVT provides a lower bound on the time stamp of any future rollback. • storage for events and state vectors older than GVT (except one state vector) can be reclaimed • I/O operations with time stamp less than GVT can be performed. • GVT algorithms are similar to LBTS algorithms in conservative synchronization Observation: The computation corresponding to GVT will not be rolled back, guaranteeing forward progress.

Time Warp and Chandy/Misra Performance • • eight processors closed queueing network, hypercube topology high priority jobs preempt service from low priority jobs (1% high priority) exponential service time (poor lookahead)

Other Optimistic Algorithms Principal goal: avoid excessive optimistic execution A variety of protocols have been proposed, among them: • window-based approaches – only execute events in a moving window (simulated time, memory) • risk-free execution – only send messages when they are guaranteed to be correct • add optimism to conservative protocols – specify “optimistic” values for lookahead • introduce additional rollbacks – triggered stochastically or by running out of memory • hybrid approaches – mix conservative and optimistic LPs • scheduling-based – discriminate against LPs rolling back too much • adaptive protocols – dynamically adjust protocol during execution as workload changes

Summary of Optimistic Algorithms Pro: • Good performance reported for a variety of applications (queueing communication networks, combat models, transportation systems) • Good transparency: offers the best hope for “general purpose” parallel simulation software (more resilient to poor lookahead than conservative methods) Con: • Memory requirements may be large • Implementation is generally more complex and difficult to debug than conservative mechanisms – System calls, memory allocation. . . – Must be able to recover from exceptions • Use in federated simulations requires adding rollback capability to existing simulations

Part II: Time Management • • Parallel discrete event simulation Conservative synchronization Optimistic synchronization Time Management in the High Level Architecture

High Level Architecture (HLA) • based on a composable “system of systems” approach – no single simulation can satisfy all user needs – support interoperability and reuse among Do. D simulations • federations of simulations (federates) – pure software simulations – human-in-the-loop simulations (virtual simulators) – live components (e. g. , instrumented weapon systems) • mandated as the standard reference architecture for all M&S in the U. S. Department of Defense (September 1996) The HLA consists of • rules that simulations (federates) must follow to achieve proper interaction during a federation execution • Object Model Template (OMT) defines the format for specifying the set of common objects used by a federation (federation object model), their attributes, and relationships among them • Interface Specification (IFSpec) provides interface to the Run-Time Infrastructure (RTI), that ties together federates during model execution

HLA Federation Interconnecting autonomous simulators Federation Simulation (federate) Interface Specification Runtime Infrastructure(RTI) Services to create and manage the execution of the federation • Federation setup / tear down • Transmitting data among federates • Synchronization (time management)

Interface Specification Category Functionality Federation Management Create and delete federation executions join and resign federation executions control checkpoint, pause, resume, restart Declaration Management Establish intent to publish and subscribe to object attributes and interactions Object Management Create and delete object instances Control attribute and interaction publication Create and delete object reflections Ownership Management Transfer ownership of object attributes Time Management Coordinate the advance of logical time and its relationship to real time Data Distribution Management Supports efficient routing of data

A Typical Federation Execution initialize federation • Create Federation Execution (Federation Mgt) • Join Federation Execution (Federation Mgt) declare objects of common interest among federates • Publish Object Class (Declaration Mgt) • Subscribe Object Class Attribute (Declaration Mgt) exchange information • Update/Reflect Attribute Values (Object Mgt) • Send/Receive Interaction (Object Mgt) • Time Advance Request, Time Advance Grant (Time Mgt) • Request Attribute Ownership Assumption (Ownership Mgt) • Modify Region (Data Distribution Mgt) terminate execution • Resign Federation Execution (Federation Mgt) • Destroy Federation Execution (Federation Mgt)

HLA Message Ordering Services The HLA provides two types of message ordering: • receive order (unordered): messages passed to federate in an arbitrary order • time stamp order (TSO): sender assigns a time stamp to message; successive messages passed to each federate have non-decreasing time stamps Property Latency reproduce before and after relationships? all federates see same ordering of events? execution repeatable? typical applications Receive Order (RO) Time Stamp Order (TSO) low higher no yes training, T&E analysis • receive order minimizes latency, does not prevent temporal anomalies • TSO prevents temporal anomalies, but has somewhat higher latency

Advancing Logical Time HLA TM services define a protocol for federates to advance logical time; logical time only advances when that federate explicitly requests an advance • Time Advance Request: time stepped federates • Next Event Request: event stepped federates • Time Advance Grant: RTI invokes to acknowledge logical time advances federate Time Advance Request or Next Event Request Time Advance Grant RTI If the logical time of a federate is T, the RTI guarantees no more TSO messages will be passed to the federate with time stamp < T Federates responsible for pacing logical time advances with wallclock time in real-time executions

HLA Time Management Services event ordering receive order messages time stamp order messages Runtime Infrastructure (RTI) FIFO queue time synchronized delivery time stamp ordered queue state updates and interactions federate logical time advances • local time and event management • mechanism to pace execution with wallclock time (if necessary) (synchronized with • federate specific techniques (e. g. , other processors) compensation for message latencies)

Synchronizing Message Delivery Goal: process all events (local and incoming messages) in time stamp order; To support this, RTI will • Deliver messages in time stamp order (TSO) • Synchronize delivery with simulation time advances TSO messages local events next TSO message RTI T’ federate current time next local event T logical time Federate: next local event has time stamp T • • If no TSO messages w/ time stamp < T, advance to T, process local event If there is a TSO message w/ time stamp T’ ≤ T, advance to T’ and process TSO message

Next Event Request (NER) • Federate invokes Next Event Request (T) to request its logical time be advanced to time stamp of next TSO message, or T, which ever is smaller • If next TSO message has time stamp T’ ≤ T – RTI delivers next TSO message, and all others with time stamp T’ – RTI issues Time Advance Grant (T’) • Else – RTI advances federate’s time to T, invokes Time Advance Grant (T) Typical execution sequences Federate RTI NER(T) Federate RAV (T’) RTI NER(T) NER: Next Event Request TAG(T) RAV (T’) Federate calls in black RTI callbacks in red TAG(T’) RTI delivers events Wall clock time TAG: Time Advance Grant RAV: Reflect Attribute Values no TSO events

Lookahead in the HLA • Each federate must declare a non-negative lookahead value • Any TSO sent by a federate must have time stamp at least the federate’s current time plus its lookahead • Lookahead can change during the execution (Modify Lookahead) – increases take effect immediately – decreased do not take effect until the federate advances its logical time T T+L Logical time 1. Current time is T, lookahead L 2. Request lookahead decrease by ∆L to L’ T+L Logical time 3. Advance ∆T, lookahead, decreases ∆T T+L Logical time 4. After advancing ∆L, lookahead is L’ L- ∆T ∆T T+ ∆T L’ ∆L T+∆L

Minimum Next Event Time and LBTS network TSO queue 13 LBTS 0=10 11 MNET 0=8 8 Current Time =8 Lookahead = 2 Current Time =9 Lookahead = 8 Federate 0 Federate 1 Federate 2 • LBTSi: Lower Bound on Time Stamp of TSO messages that could later be placed into the TSO queue for federate i – TSO messages w/ TS ≤ LBTSi eligible for delivery – RTI ensures logical time of federate i never exceeds LBTSi • MNETi: Minimum Next Event Time is a lower bound on the time stamp of any message that could later be delivered to federate i. – Minimum of LBTSi and minimum time stamp of messages in TSO queue

Event Retraction Previously sent events can be “unsent” via the Retract service – Update Attribute Values and Send Interaction return a “handle” to the scheduled event – Handle can be used to Retract (unschedule) the event – Can only retract event if its time stamp > current time + lookahead – Retracted event never delivered to destination (unless Flush Queue used) Sample execution sequence: Vehicle 1. NER (100) 2. Receive Interaction (90) 1. Handle=UAV (100) 2. TAG (90) NER: Next Event Request UAV: Update Attribute Values) TAG: Time Advance Grant (callback) 3. Retract(Handle) Observer Wallclock time 1. Vehicle schedules position update at time 100, ready to advance to time 100 2. receives interaction (break down event) invalidating position update at time 100 3. Vehicle retracts update scheduled for time 100

Optimistic Time Management Mechanisms to ensure events are processed in time stamp order: • conservative: block to avoid out of order event processing • optimistic: detect out-of-order event processing, recover (e. g. , Time Warp) Primitives for optimistic time management • Optimistic event processing – Deliver (and process) events without time stamp order delivery guarantee – HLA: Flush Queue Request • Rollback – Deliver message w/ time stamp T, other computations already performed at times > T – Must roll back (undo) computations at logical times > T – HLA: (local) rollback mechanism must be implemented within the federate • Anti-messages & secondary rollbacks – Anti-message: message sent to cancel (undo) a previously sent message – Causes rollback at destination if cancelled message already processed – HLA: Retract service; deliver retract request if cancelled message already delivered • Global Virtual Time – Lower bound on future rollback to commit I/O operations, reclaim memory – HLA: Query Next Event Time service gives current value of GVT

Optimistic Time Management in the HLA Support for Optimistic Federates • federations may include conservative and/or optimistic federates • federates not aware of local time management mechanism of other federates (optimistic or conservative) • optimistic events (events that may be later canceled) will not be delivered to conservative federates that cannot roll back • optimistic events can be delivered to other optimistic federates • individual federates may be sequential or parallel simulations Flush Queue Request: similar to NER except (1) deliver all messages in RTI’s local message queues, (2) need not wait for other federates before issuing a Time Advance Grant

Summary: HLA Time Management Functionality: • allows federates with different time management requirements (and local TM mechanisms) to be combined within a single federation execution – DIS-style training simulations – simulations with hard real-time constraints – event-driven simulations – time-stepped simulations – optimistic simulations HLA Time Management services: • Event order – receive order delivery – time stamp order delivery • Logical time advance mechanisms – TAR/TARA: unconditional time advance – NER/NERA: advance depending on message time stamps

Part III: Distributed Virtual Environments • Introduction • Dead reckoning • Data distribution

Example: Distributed Interactive Simulation (DIS) “The primary mission of DIS is to define an infrastructure for linking simulations of various types at multiple locations to create realistic, complex, virtual ‘worlds’ for the simulation of highly interactive activities” [DIS Vision, 1994]. • developed in U. S. Department of Defense, initially for training • distributed virtual environments widely used in Do. D; growing use in other areas (entertainment, emergency planning, air traffic control)

A Typical DVE Node Simulator network visual display system Image Generator terrain database control/ display interface Other Vehicle State Table own vehicle dynamics network interface sound generator controls and panels Execute every 1/30 th of a second: • receive incoming messages & user inputs, update state of remote vehicles • update local display • for each local vehicle • compute (integrate) new state over current time period • send messages (e. g. , broadcast) indicating new state Reproduced from Miller, Thorpe (1995), “SIMNET: The Advent of Simulator Networking, ” Proceedings of the IEEE, 83(8): 1114 -1123.

Typical Sequence visual display 3 6 7 4 8 1. Detect trigger press Image Generator terrain database control/ display interface Other Vehicle State Table own vehicle dynamics 1 network interface 2. Audio “fire” sound 3. Display muzzel flash 4. Send fire PDU sound generator 5. Display muzzel flash 2 6. Compute trajectory, display tracer 7. Display shell impact Controls/panels visual display 5 9 Image Generator terrain database control/ display interface 4 Other Vehicle State Table own vehicle dynamics 10 Controls/panels network interface sound generator 8 11 8. Send detonation PDU 9. Display shell impact 10. Compute damage 11. Send Entity state PDU indicating damage

DIS Design Principles • • • Autonomy of simulation nodes – simulations broadcast events of interest to other simulations; need not determine which others need information – receivers determine if information is relevant to it, and model local effects of new information – simulations may join or leave exercises in progress Transmission of “ground truth” information – each simulation transmits absolute truth about state of its objects – receiver is responsible for appropriately “degrading” information (e. g. , due to environment, sensor characteristics) Transmission of state change information only – if behavior “stays the same” (e. g. , straight and level flight), state updates drop to a predetermined rate (e. g. , every five seconds) “Dead Reckoning” algorithms – extrapolate current position of moving objects based on last reported position Simulation time constraints – many simulations are human-in-the-loop – humans cannot distinguish temporal difference < 100 milliseconds – places constraints on communication latency of simulation platform

Part III: Distributed Virtual Environments • Introduction • Dead reckoning • Data distribution

Distributed Simulation Example • Virtual environment simulation containing two moving vehicles • One vehicle per federate (simulator) • Each vehicle simulator must track location of other vehicle and produce local display (as seen from the local vehicle) • Approach 1: Every 1/60 th of a second: – Each vehicle sends a message to other vehicle indicating its current position – Each vehicle receives message from other vehicle, updates its local display

Limitations • Position information corresponds to location when the message was sent; doesn’t take into account delays in sending message over the network • Requires generating many messages if there are many vehicles

Dead Reckoning • Send position update messages less frequently • local dead reckoning model predicts the position of remote entities between updates visual display system “infrequent” position update messages get position at frame rate last reported state: position = (1000, 1000) traveling east @ 50 feet/sec (1050, 1000) 1000 Image Dead reckoning Generator model predicted position one second later terrain database 1000 • When are updates sent? • How does the DRM predict vehicle position? 1050

Re-synchronizing the DRM When are position update messages generated? • Compare DRM position with exact position, and generate an update message if error is too large • Generate updates at some minimum rate, e. g. , 5 seconds (heart beats) High Fidelity Model aircraft 1 close enough? timeout? DRM aircraft 1 simulator for aircraft 1 DRM aircraft 2 over threshold or timeout entity state update PDU DRM aircraft 1 display sample DRM at frame rate simulator for aircraft 2

Dead Reckoning Models • • P(t) = precise position of entity at time t Position update messages: P(t 1), P(t 2), P(t 3) … v(ti), a(ti) = ith velocity, acceleration update DRM: estimate D(t), position at time t – ti = time of last update preceding t – ∆t = ti - t • Zeroth order DRM: – D(t) = P(ti) • First order DRM: – D(t) = P(ti) + v(ti)*∆t • Second order DRM: – D(t) = P(ti) + v(ti)*∆t + 0. 5*a(ti)*(∆t)2

DRM Example DRM estimates position A true position state update message DRM estimate of true position B C t 1 t 2 generate state update message receive message just before screen update D E display update Potential problems: • Discontinuity may occur when position update arrives; may produce “jumps” in display • Does not take into account message latency

Time Compensation Taking into account message latency • Add time stamp to message when update is generated (sender time stamp) • Dead reckon based on message time stamp A true position state update message DRM estimate of true position display update B C t 1 t 2 update with time compensation D E

Smoothing Reduce discontinuities after updates occur • “phase in” position updates • After update arrives – Use DRM to project next k positions – Interpolate position of next update A B C true position state update message interpolated position t 1 t 2 D DRM estimate of true position display update E extrapolated position used for smoothing Accuracy is reduced to create a more natural display

Part III: Distributed Virtual Environments • Introduction • Dead reckoning • Data distribution

Data Distribution Management A federate sending a message (e. g. , updating its current location) cannot be expected to explicitly indicate which other federates should receive the message Basic Problem: which federates receive messages? • broadcast update to all federates (Sim. Net, early versions of DIS) – does not scale to large numbers of federates • grid sectors – OK for distribution based on spacial proximity • routing spaces (STOW-RTI, HLA) – generalization of grid sector idea Content-based addressing: in general, federates must specify: • Name space: vocabulary used to specify what data is of interest and to describe the data contained in each message (HLA: routing space) • Interest expressions: indicate what information a federate wishes to receive (subset of name space; HLA: subscription region) • Data description expression: characterizes data contained within each message (subset of name space; HLA: publication region)

HLA Routing Spaces 1. 0 • Federate 1 (sensor): subscribe to S 1 • Federate 2 (sensor): subscribe to S 2 • Federate 3 (target): update region U S 2 0. 5 0. 0 S 1 U update messages by target are sent to federate 1, but not to federate 2 0. 5 1. 0 HLA Data Distribution Management • N dimensional normalized routing space • interest expressions are regions in routing space (S 1=[0. 1, 0. 5], [0. 2, 0. 5]) • each update associated with an update region in routing space • a federate receives a message if its subscription region overlaps with the update region

Implementation: Grid Based 1. 0 21 22 23 24 25 16 17 18 19 20 13 14 15 S 2 0. 5 0. 0 11 S 1 12 U 6 7 8 9 10 1 2 3 4 5 0. 5 partition routing space into grid cells, map each cell to a multicast group, id=15 U publishes to 12, 13 S 1 subscribes to 6, 7, 8, 11, 12, 13 S 2 subscribes to 11, 12, 16, 17 1. 0 • subscription region: Join each group overlapping subscription region • attribute update: send Update to each group overlapping update region • need additional filtering to avoid unwanted messages, duplicates

Changing a Subscription Region 33 34 35 36 37 38 39 40 25 26 27 28 29 30 31 32 17 18 19 20 21 22 23 24 Join group 9 10 11 12 13 14 15 16 no operations issued 1 2 3 4 5 6 7 8 new region existing subscription region Leave group Change subscription region: • issue Leave operations for (cells in old region - cells in new region) • issue Join operations for (cells in new region - cells in old region) Observation: each processor can autonomously join/leave groups whenever its subscription region changes w/o communication.

Another Approach: Region-Based Implementation 1. 0 • • • S 2 0. 5 0. 0 S 1 U 0. 5 Multicast group associated with U Membership of U: S 1 Modify U to U’: – determine subscription regions overlapping with U’ – Modify membership accordingly 1. 0 • Associate a multicast group with each update region • Membership: all subscription regions overlapping with update region • Changing subscription (update) region – Must match new subscription (update) region against all existing update (subscription) regions to determine new membership • No duplicate or extra messages • Change subscription: typically requires interprocessor communication • Can use grids (in addition to regions) to reduce matching cost

Distributed Virtual Environments: Summary • Perhaps the most dominant application of distributed simulation technology to date – Human in the loop: training, interactive, multi-player video games – Hardware in the loop • Managing interprocessor communication is the key – Dead reckoning techniques – Data distribution management • Real-time execution essential • Many other issues – – Terrain databases, consistent, dynamic terrain Real-time modeling and display of physical phenomena Synchronization of hardware clocks Human factors

Future Research Directions

The good news. . . Parallel and distributed simulation technology is seeing widespread use in real-world applications and systems • HLA, standardization (OMG, IEEE 1516) • Other defense systems • Interactive games (no time management yet, though!)

Interoperability and Reuse the principal reason distributed simulation technology is seeing widespread use (scalability still important!) “Sub-objective 1 -1: Establish a common high level simulation architecture to facilitate interoperability [and] reuse of M&S components. ” - U. S. Do. D Modeling and Simulation Master Plan “critical capability that underpin the IMTR vision [is]. . . total, seamless model interoperability: Future models and simulations will be transparently compatible, able to plugand-play …” - Integrated Manufacturing Technology Roadmap • Enterprise Management • Circuit design, embedded systems • Transportation • Telecommunications

Research Challenges • Shallow (communication) vs. deep (semantics) interoperability: How should I develop simulations today, knowing they may be used tomorrow for entirely different purposes? • Multi-resolution modelling • Smart, adaptable interfaces? • Time management: A lot of work completed already, but still not a solved problem! – Conservative: zero lookahead simulations will continue to be the common case – Optimistic: automated optimistic-ization of existing simulations (e. g. , state saving) – Challenge basic assumptions used in the past: New problems & solutions, not new solutions to the old problems • Relationship of interoperability and reuse in simulation to other domains (e. g. , e-commerce)?

Real-Time Systems Interactive distributed simulations for immersive virtual environments represents fertile new ground for distributed simulation research The distinction between the virtual and real world is disappearing

Research Challenges • Real-time management – – Time managed, real-time distributed simulations? Predictable performance Performance tunable (e. g. , synchronization overheads) Mixed time management (e. g. , federations with both analytic and real-time training simulations) • Distributed simulation over WANs, e. g. , the Internet – Server vs. distributed architectures? – PDES over unreliable transport and nodes • Distributed simulations that work! – High quality, robust systems • Data Distribution Management – – Implementation architectures Managing large numbers of multicast groups Join/leave times Integration and exploitation of network Qo. S

Simulation-Based Decision Support Real World System Real-Time Advisor Automated and/or interactive analysis tools analysts live data feeds Information system forecasting tool (fast simulation) • Battle management • Enterprise control • Transportation Systems decision makers

Research Challenges • Ultra fast model execution – Much, much faster-than-real-time execution – Instantaneous model execution: Spreadsheet-like performance • Ultra fast execution of multiple runs • Automated simulation analyses: smart, self-managed devices and systems? • Ubiquitous simulation?

Closing Remarks After years of being largely an academic endeavor, parallel and distributed simulation technology is beginning to thrive. Many major challenges remain to be addressed before the technology can achieve its fullest potential.

The End