Experimental Results to the Paper “Parallel Processing for Business Artifacts with Declarative Lifecycles”

This paper serves as an appendix to [4]. Two variants of the parallel algorithms are provided. Experiments upon sequential and parallel algorithms are conducted and analyzed. 1 Two Variants of the Parallel Algorithms This section introduces two variants of the parallel algorithms shown in Section 5 of [4], namely “event-based parallel algorithm” (which is the same as the one in Section 7 of [4]) and “node-based parallel algorithm”. These two algorithms all follow the idea in [4] but different in implementation level. In detail, these two algorithms have different ways to allocate threads for parallel execution. Event-based parallel algorithm For the event-based parallel algorithm, we propose a new role, called “event thread”, which will be created for every incoming event. The event thread is responsible for selecting the available nodes (similar to a scheduler) and for evaluating all the reachable nodes for its associated incoming event together with updating the corresponding snapshot (similar to an evaluation scheduler). In essence, the improved parallel algorithm “projects” a centralized scheduler down to each event thread so that the event thread can evaluate a node as soon as it is available, and it works as follows: 1. When an incoming event arrives, the system allocates an event thread for it. 2. The event thread first atomically labels all the reachable nodes for the current event with the sequence number assigned (similar to lines 9 11 of Alg. 1 in [4]). 3. Then the event thread traverses the reachable nodes for the current incoming event according to the topological order. 4. During the traversal, if the node to be visited satisfies the evaluation policy, then the event thread will evaluate the node and update the corresponding snapshot in the parallel evaluation matrix; otherwise, the thread will wait till the node becomes available. 5. When traversal finishes, the thread updates the blank attributes in the corresponding snapshot. ? This paper appears as [4] Node-based parallel algorithm Similar to the event-based parallel algorithm, the node-based parallel algorithm allocates several threads to evaluate the sentries in parallel. Instead of creating a thread for each incoming event, the node-based parallel algorithm creates a thread, called “node thread” for each node in the EPDG. The algorithm works as follows: 1. Once the EPDG is loaded into the memory, the system allocates a thread for each node. 2. Once an incoming event arrives, the system labels all the reachable nodes according to the topological order. 3. For each node thread, it evaluates the corresponding sentries if the smallest number in its will visit set satisfies the evaluation policy. 4. Each node thread will keep evaluating until the EPDG is unloaded or there is no more events labeled to its will visit set. One advantage of the node-based parallel algorithm is to realize the parallelism even within a B-step. When an event needs to traverse all its reachable nodes and there is a branch in its reachable subgraph, then the node threads of the two branches can work separately in parallel. The analysis of the node-based parallel algorithm is not straight-forward, thus it is evaluated in the next section by experiments.