Parallel Discrete Event Simulation (PDES)

PDES advances simulation time by jumping directly from one event timestamp to the next, skipping idle periods. Each event represents a state change—such as an order arrival or machine failure—and is processed as an instantaneous occurrence. This contrasts with time-stepped simulation, which wastes compute cycles on intervals where nothing changes. Event granularity directly impacts performance: fine-grained events increase parallelism overhead, while coarse-grained events reduce potential concurrency.

A protocol that prevents causality errors by blocking a logical process (LP) from executing an event until it can guarantee no straggler event with a smaller timestamp will arrive. The Chandy-Misra-Bryant algorithm uses null messages carrying timestamp promises to avoid deadlock. Conservative protocols excel when events are densely connected across LPs, but can suffer from low processor utilization if lookahead—the minimum future time an LP can predict—is small.

LPs execute events speculatively without waiting for guarantees, detecting causality violations after the fact. The Time Warp mechanism rolls back erroneous computations using anti-messages and restores prior state via checkpointing. Optimistic protocols exploit parallelism aggressively and adapt well to dynamic workloads, but require careful memory management to bound the overhead of state saving and the cost of cascading rollbacks.

A computed lower bound on the timestamp of any event that could ever be rolled back in an optimistic simulation. GVT defines the commit horizon: events with timestamps below GVT are permanent and their effects can be released to external systems. It also governs fossil collection—the reclamation of memory consumed by old state vectors and event histories. Efficient GVT estimation algorithms, such as Mattern's method, are critical for sustained performance.

The ability of an LP to predict a lower bound on the timestamp of future events it will generate. A large lookahead value allows conservative protocols to safely advance simulation time and enables optimistic protocols to reduce rollback probability. Lookahead is derived from model semantics: a minimum service time, a transmission delay, or a contractual lead time. Exploiting lookahead is the single most effective optimization in PDES performance tuning.

The decomposition of a simulation model into logical processes (LPs) that communicate exclusively through timestamped event messages. Effective partitioning minimizes cross-LP communication while balancing computational load. Common strategies include spatial partitioning (assigning geographic regions to processors), functional partitioning (separating distinct activities like manufacturing and shipping), and scatter-gather patterns for embarrassingly parallel Monte Carlo trials.

A protocol that strictly prevents causality errors by only processing an event when it is safe to do so. A logical process (LP) blocks execution until it can guarantee no straggler event with a smaller timestamp will arrive from another LP.

• Chandy-Misra-Bryant algorithm: The classic protocol using null messages to propagate lower-bound timestamps
• Lookahead: The minimum simulation time into the future an LP can predict its next outgoing event—critical for performance
• Deadlock avoidance: Null messages circulate to break circular wait conditions without requiring global detection

LPs process events aggressively without waiting for safety guarantees, detecting and recovering from causality violations via rollback mechanisms. The Time Warp protocol is the canonical implementation.

• Anti-messages: Sent to annihilate previously transmitted messages that are now invalidated by a rollback
• State saving: Periodic checkpointing of LP state vectors to enable efficient restoration
• Fossil collection: Reclaiming memory for states and events older than Global Virtual Time (GVT)

A monotonically increasing lower bound on the timestamp of any event that could ever be rolled back in an optimistic simulation. GVT defines the commit horizon.

• Commit operations: I/O and external interactions are held until their timestamp is less than GVT
• Memory management: States and events older than GVT are eligible for fossil collection
• Computation: Typically calculated via distributed reduction across all LPs, finding the minimum of local virtual times and in-transit message timestamps

The decomposition of a simulation model into logical processes (LPs) that can execute concurrently on separate processors. Partitioning quality directly determines speedup.

• Spatial partitioning: Dividing the simulated topology geographically (e.g., warehouse zones, network segments)
• Temporal load balancing: Redistributing LPs dynamically to mitigate stragglers and idle processors
• Communication overhead: The cost of inter-LP event messages often dominates, making locality-preserving partitions critical

The mechanism by which an optimistic simulator identifies that events have been processed out of timestamp order, violating the local causality constraint.

• Straggler detection: An incoming event with a timestamp earlier than the LP's local virtual time triggers a rollback
• Cascading rollbacks: A single causality error can propagate anti-messages to other LPs, potentially causing secondary rollbacks
• Bounded optimism: Hybrid approaches that throttle speculative execution to limit the depth and cost of rollback cascades