Generalizing Bulk-Synchronous Parallel Processing for Data Science: From Data to Threads and Agent-Based Simulations

Abstract

We generalize the bulk-synchronous parallel (BSP) processing model to make it better support agent-based simulations. Such simulations frequently exhibit hierarchical structure in their communication patterns which can be exploited to improve performance. We allow for the creation of temporary artificial network partitions during which agents synchronize only locally within their group in a way that does not compromise the correctness of a simulation. We have built a distributed engine, CloudCity, which uses this idea to improve the locality of computation, communication, and synchronization in such simulations. We experimentally evaluate the performance of our system on a benchmark of simulation workloads and compare it against other popular BSP-like systems, obtaining insights into the impact of various system design choices and optimization on simulation engine performance.

Multi-Stage Vertex-Centric Programming for Agent-Based Simulations

“…A natural question is that given all existing graph processing frameworks that already support vertex-centric programming, why is there a need for another framework for agent-based simulations? This question is addressed in [35], where our experiments showed that the performance of these existing frameworks could differ by up to three orders of magnitude when executing a benchmark consisting of representative agent-based simulation workloads selected from population dynamics, economics, and epidemics, due to different design choices of these systems.…”

“…We implement the staging-time wait() instruction in Cloud-City [35], a distributed agent-based simulation system. An agent is single-threaded and communicates by sending messages.…”

“…4 Just like coroutines, computations of a vertex function yield when executing wait() and later resume from the last saved execution point when the next superstep begins, avoiding the interpretation overhead caused by dispatching on the value of a superstep. The previous pseudocode can be expressed as follows using wait(): We implement wait() in the programming model of a distributed agent-based simulation system CloudCity [35] and evaluate the performance benefit of staging compared with unstaged vertex programs. Our results show that for an agent with 5 to 20 branches, staging reduces the total execution time by 10% -25%.…”

“…Compare with[35], we eliminate 𝑑𝑒𝑙𝑎𝑦 from the signature of RPC instructions callAndForget and asyncCall for simplicity 8. For simplicity, here we consider wait( ), as we have discussed in the previous section.…”

“…For simplicity, here we consider wait( ), as we have discussed in the previous section. This is different from wait(n) : Unit in[35], which takes an integer parameter. The semantics of wait( ) is the same as wait(1).…”