Measuring the effectiveness of task-level parallelism for high-level vision

In this paper, we focus on task-level parallelism, which is obtained by a high-level decomposition of the production system. Speed-ups obtained from task-level parallelism will multiply with che speed-ups obtained from match parallelism. Our vehicle for the investigation of task-level parallelism is SPAM, a high-level vision system, implemented in a production system architecture. SPAM is a mature research system having over 600 productions, with a typical scene analysis task having between 50,000 to 400,000 production firings and an execution time of the order of 10 to 100 cpu hours. We present a characterization of task-level parallelism in production systems and, from that, select an explicit, data-driven approach for exploiting task-level parallelism. We describe a methodology for applying the chosen approach to obtain a parallel task decomposition of SPAM and to arrive at our parallel implementation, SPAM/PSM. We present the results of that implementation that show near linear speed-ups of over 12 fold using 14 processors and that point the way to substantial speedups from task-level parallelism. 1. I n t r o d u c t i o n Large production systems (rule-based systems) continue to suffer from extremely slow execution which limits their utility in practical applications as well as research settings. Most efforts at speeding up these systems have focused on match, i.e., knowledge-search, parallelism in production systems [3, 5, 7, 15, 20, 21]. Though good speed-ups have been achieved in this process, the total speed-up available from this source is limited. Therefore, match parallelism alone will not alleviate the problem of slow execution in production systems. In this paper, we focus on task-level parallelism, which is obtained by a high-level decomposition of the production system. Speed-ups obtained from task-level parallelism will multiply with the speed-ups obtained from match parallelism. Our vehicle for the investigation of task-level parallelism is SPAM [12, 13, 14], a high-level vision system, implemented in a production system architecture. SPAM is a mature research system having over 600 productions, with a typical scene analysis task requiring between 50,000 to 400,000 production firings and an execution time of the order of 10 to 100 cpu hours 1 . Unlike most other production systems examined for studies in parallelism, it has embedded in it a large computational demand related to the vision task that it performs. This task-related computation is separate from the computation performed for knowledgesearch in the system. This is evident in the large RHS processing time for this system. While many production systems spend up to 90% of their time in knowledge-search, SPAM spends only about 30-50% of its time there. l T h e s e measurements are taken from the Lisp-based version of OPS5 running on a VAX/785 processor. MEASURING THE EFFECTIVENESS OF TASK-LEVEL PARALLELISM FOR HIGH-LEVEL VISION 2 In this paper, we show that the opportunities for task-level parallelism in SPAM are high and provide a much larger payoff in speed-up than match parallelism. We present a methodology and a set of principles to arrive at a suitable parallel decomposition of the SPAM task that results in near linear speed-ups of over 12 fold using 14 processors on a 16-processor shared-memory multiprocessor. Our results also indicate that a potential speed-up of 50 to 100 fold may be achievable due to task-level parallelism. We further show that match parallelism, when used in conjunction with task-level parallelism, gives another multiplicative factor of speed-up which is proportional to the size of the match component in the overall execution time. In the SPAM system, this additional multiplicative factor is around 1.5 to 2. This paper is organized as follows: Section 2 provides some background about production systems and SPAM, the image interpretation system that is the focus of our analysis of task-level parallelism. Section 3 discusses match parallelism and task-level parallelism in production systems. We describe a new organization to compare previous work in task-level parallelism along several independent dimensions. Section 4 discusses the implementation methodology used to determine appropriate levels for task-level parallelism. We also describe a set of experiments and measurements on SPAM that allowed us to select an appropriate grain of decomposition. These techniques should be applicable to the analysis of other large production systems for evaluating the opportunities for tasklevel parallelism. A new system, SPAM/PSM, resulted from the application of this methodology and its implementation is described in Section 5. Section 6 presents a detailed analysis of the results of experiments across several dimensions including grain of decomposition, speed-ups due to processor allocation for match-level and task-level parallelism. Finally, Section 7 presents a summary of our research results and Section 8 discusses some issues for future work. 2. Background In this section we provide a brief overview of OPS5 and SPAM. SPAM is implemented in OPS5, hence the description of OPS5 will be useful in understanding some of the issues in how SPAM represents knowledge about spatial and structural constraints used in computer vision. Besides providing background information, this section introduces the terminology that will be used in the rest of this paper. 2.1. OPS5 An OPS 5 [2] production system is composed of a set of if-then rules, called productions, that make up the production memory, and a database of temporary data structures, called the working memory. The individual data structures are called working memory elements (WMEs), and are lists of attribute-value pairs. Each production consists of a conjunction of condition elements (CEs) corresponding to the if pan of the rule (also called the left-hand side or LHS), and a set of actions corresponding to the then part of the rule (also called the right-hand side or RHS). The CEs in a production consist of attribute-value tests, where some attributes may contain variables as values. The attribute-value tests of a CE must all be matched by a WME for the CE to match; the variables in the condition element may match any value, but if the variable occurs in more than one CE of a production, then all occurrences of the variable must match identical values. When all the CEs of a production are matched, the production is satisfied, and an instantiation of the production (a list of WMEs that matched it), is created and entered into the conflict set. The MEASURING THE EFFECTIVENESS OF TASK-LEVEL PARALLELISM FOR HIGH-LEVEL VISION 3 production system uses a selection procedure called conflict-resolution to choose a production from the conflict set, which is then fired. When a production fires, the RHS actions associated with that production are executed. The RHS actions can add, remove or modify WMEs, or perform I/O. The production system is executed by an interpreter that repeatedly cycles through three steps: 1. Match 2. Conflict-resolution 3. Act The matching procedure determines the set of satisfied productions, the conflict-resolution procedure selects a single instantiation, and the act procedure executes its RHS. These three steps are collectively called the recognize-act cycle. 2.2. SPAM: A Production System Architecture For Scene Interpretation SPAM [12, 13, 11] is a production system architecture for the interpretation of aerial imagery with applications to automated cartography and digital mapping. It tests the hypothesis that the interpretation of aerial imagery requires substantial knowledge about the scene under consideration. Knowledge about the type of scene — airport, suburban housing development, urban city — aids in low-level and intermediate level image analysis, and will drive high-level interpretation by constraining search for plausible consistent scene models. SPAM has been applied in two task areas: airport and suburban house scene analysis. In the remainder of this section we describe the SPAM architecture, and give run-time statistics that lead us to focus on one phase for our studies in parallelism.