Using machine learning to optimize graph execution on NUMA machines

Improving the efficiency of graph algorithm executions on high‐performance computing

“…Traditionally, HPC servers receive batches of graphs applications to execute serially (i.e., one after another), always using all the available processing resources ‐ in the same way as any other HPC application. However, graph applications, because of their irregular structure and larger dimension, are intrinsically communication/memory‐bound, which means that they tend to use the CPU less when compared to the average parallel application 1,10 . Given that, graph applications executions are even more affected by issues that tend to limit the scalability of parallel applications related to both hardware (e.g., saturation of execution units and communication bus) and software (e.g., data synchronization and concurrent accesses to shared memory) 11‐13 .…”

“…The rising number of cores and available memory in high-performance computing (HPC) servers has enabled the analysis of a massive amount of graph-structured data extracted from sources like Google, Facebook, and Twitter. [1][2][3] Simultaneously, the unprecedented growth of such interconnected data has been pushing forward the development of efficient graph analytics methods to extract useful information from massive data sources, making enhancements in several areas, such as business, geolocation, fraud detection, and social network analysis. 4,5 Graph algorithms such as PageRank and single source shortest-paths (SSSP) make it possible to perform operations like the page ordering of the most visited content on Google or even complex computations in the Artificial Intelligence ambit.…”

“…However, graph applications, because of their irregular structure and larger dimension, are intrinsically communication/memory-bound, which means that they tend to use the CPU less when compared to the average parallel application. 1,10 Given that, graph applications executions are even more affected by issues that tend to limit the scalability of parallel applications related to both hardware (e.g., saturation of execution units and communication bus) and software (e.g., data synchronization and concurrent accesses to shared memory). [11][12][13] This added to the fact that HPC multiprocessors have their memory hierarchy getting more complex, may indicate that executing all graph applications with the maximum number of threads, which is the default behavior, may not be efficient.…”

Interference Analysis and Resource Management in Server Processors: from HPC to Cloud Computing

“…These approaches, however, provide little flexibility as they require knowing the running applications before the execution starts and do not allow adapting the number of spawned threads dynamically. PredG [67] uses machine learning to select the best thread and data mapping policies to run graph applications on a NUMA system. Both approaches require knowing application-level information, such as the input graphs for decision-making.…”