Much of the data being produced in large scale by modern applications represents connected entities and their relationships, that can be modeled as large graphs. In order to extract valuable information from these large datasets, several parallel and distributed graph processing engines have been proposed. These systems are designed to run in large clusters, where resources must by allocated efficiently. Aiming to handle this problem, this paper presents a performance prediction model for GPS, a popular Pregel-based graph processing framework. By leveraging a micro-partitioning technique, our system can use various partitioning algorithms that greatly reduce the execution time, comparing with the simple hash partitioning that is commonly used in graph processing systems. Experimental results show that the prediction model has accuracy close to 90%, allowing it to be used in schedulers or to estimate the cost of running graph processing tasks.

现代应用程序大规模生成的大部分数据表示连接的实体及其关系，可以将其建模为大图。为了从这些大型数据集中提取有价值的信息，已经提出了几种并行和分布式图形处理引擎。这些系统被设计为在大型集群中运行，必须在其中有效分配资源。为了解决这个问题，本文提出了 GPS 的性能预测模型，这是一种流行的基于 Pregel 的图形处理框架。通过利用微分区技术，我们的系统可以使用各种分区算法，与图形处理系统中常用的简单散列分区相比，大大减少了执行时间。实验结果表明，该预测模型准确率接近90%，