A Visual Analytics System for Optimizing Communications in Massively Parallel Applications

Abstract

Current and future supercomputers have tens of thousands of compute nodes interconnected with high-dimensional networks and complex network topologies for improved performance. Application developers are required to write scalable parallel programs in order to achieve high throughput on these machines. Application performance is largely determined by efficient inter-process communication. A common way to analyze and optimize performance is through profiling parallel codes to identify communication bottlenecks. However, understanding gigabytes of profile data is not a trivial task. In this paper, we present a visual analytics system for identifying the scalability bottlenecks and improving the communication efficiency of massively parallel applications. Visualization methods used in this system are designed to comprehend large-scale and varied communication patterns on thousands of nodes in complex networks such as the 5D torus and the dragonfly. We also present efficient rerouting and remapping algorithms that can be coupled with our interactive visual analytics design for performance optimization. We demonstrate the utility of our system with several case studies using three benchmark applications on two leading supercomputers. The mapping suggestion from our system led to 38% improvement in hop-bytes for MiniAMR application on 4,096 MPI processes.

Ferromagnetic CoSe broadband photodetector at room temperature

“…( 2)) is a quadratic function, and the constraints are linear, the model can be classified as an Integer Quadratic Programming model. In fact, the problem presented is a typical Quadratic Assignment Problem (QAP [17,37]). QAP is considered to be NP-hard, however, considering the relatively small number of topics in the conversation history, this study employs the Gurobi solver to address the aforementioned problem.…”

A Visual Analytics Framework for Reviewing Streaming Performance Data

“…Similarly, Bhatele et al [5] analyzed Dragonfly-based networks using a radial layout and a matrix view to show inter-group and intra-group links between the compute nodes. Fujiwara et al [17] utilized node-link diagrams and the matrix-based representations with hierarchical aggregation techniques to generalize to any network topology. Li et al [29] developed flexible visualization to analyze the network performance on the Dragonfly network by applying data aggregation techniques to scale for large scale networks.…”

“…Because streaming performance data is continuously changing with high-volume and high-variety properties, without any algorithmic and visual supports, conducting the above analysis is almost infeasible. While existing visual analytic systems [17,29] aim to support the analysis of dynamic performance data, support for real-time analysis is lacking.…”

Performance optimality or reproducibility

“…Predicting user estimates of the runtime of their jobs so as to ensure that they not killed prematurely due to underestimation has also been studied from the point of view of noise and variation [23,66]. Furthermore, several machine learning techniques to predict application performance and detect anomalies for both power and network related variations have been explored [5,9,16,32,43,44,70,73], and the visualization community has come up with novel mechanisms to analyze network traffic and performance data [10,24,31,39,45,69]. However, all the studies listed above have focused on a single objective such as network traffic, or a single constraint such as power.…”

“…For our dataset, we varied core count (16,20,24 cores per node), task count (512, 1024, 2048, 4096 MPI ranks), power cap (64 W, 80 W, 115 W per socket), network QoS levels (combinations ranging from service level 0 to 2), and application placement algorithms (packed, spread and random assignment). All benchmarks are MPIbased, compiled with Intel compiler 16.0.3 and MVAPICH 2.2.…”