Florina M. Ciorba

Ehsan Yousefzadeh-Asl-Miandoab, Reza Karimzadeh, Bulat Ibragimov et al.

GPUs running deep learning (DL) workloads are frequently underutilized. Collocating multiple DL training tasks on the same GPU can improve utilization but introduces two key risks: (1) out-of-memory (OOM) crashes for newly scheduled tasks, and (2) severe performance interference among co-running tasks, which can negate any throughput gains. These issues reduce system robustness, quality of service, and energy efficiency. We present CARMA, a task-level, collocation-aware resource management system for the server-scale. CARMA addresses collocation challenges via (1) fine-grained monitoring and bookkeeping of GPUs and a collocation risk analysis that filters out the high-risk GPUs; (2) task placement policies that cap GPU utilization to avoid OOMs and limit interference; (3) integration of GPU memory need estimators for DL tasks to minimize OOMs during collocation; and (4) a lightweight recovery method that relaunches jobs crashed due to OOMs. Our evaluation on a DL training workload derived from real-world traces shows that CARMA uses GPUs more efficiently by making more informed collocation decisions: for the best-performing collocation policy, CARMA increases GPU streaming multiprocessor (SM) utilization by 54%, the parallelism achieved per SM by 61%, and memory use by 62%. This results in a $\sim$35% and $\sim$15% reduction in the end-to-end execution time (makespan) and GPU energy consumption, respectively, for this workload.

Jonas H. Müller Korndörfer, Ali Mohammed, Ahmed Eleliemy et al.

Scientific and data science applications are becoming increasingly complex, with growing computational and memory demands. Modern high performance computing (HPC) systems provide high parallelism and heterogeneity across nodes, devices, and cores. To achieve good performance, effective scheduling and load balancing techniques are essential. Parallel programming frameworks such as OpenMP now offer a variety of advanced scheduling algorithms to support diverse applications and platforms. This creates an instance of the scheduling algorithm selection problem, which involves identifying the most suitable algorithm for a given combination of workload and system characteristics. In this work, we explore learning-based approaches for selecting scheduling algorithms in OpenMP. We propose and evaluate expert-based and reinforcement learning (RL)-based methods, and conduct a detailed performance analysis across six applications and three systems. Our results show that RL methods are capable of learning high-performing scheduling decisions, although they require significant exploration, with the choice of reward function playing a key role. Expert-based methods, in contrast, rely on prior knowledge and involve less exploration, though they may not always identify the optimal algorithm for a specific application-system pair. By combining expert knowledge with RL-based learning, we achieve improved performance and greater adaptability. Overall, this work demonstrates that dynamic selection of scheduling algorithms during execution is both viable and beneficial for OpenMP applications. The approach can also be extended to MPI-based programs, enabling optimization of scheduling decisions across multiple levels of parallelism.

Vivek Kale, Christian Iwainsky, Michael Klemm et al.

Parallel loops are an important part of OpenMP programs. Efficient scheduling of parallel loops can improve performance of the programs. The current OpenMP specification only offers three options for loop scheduling, which are insufficient in certain instances. Given the large number of other possible scheduling strategies, it is infeasible to standardize each one. A more viable approach is to extend the OpenMP standard to allow for users to define loop scheduling strategies. The approach will enable standard-compliant application-specific scheduling. This work analyzes the principal components required by user-defined scheduling and proposes two competing interfaces as candidates for the OpenMP standard. We conceptually compare the two proposed interfaces with respect to the three host languages of OpenMP, i.e., C, C++, and Fortran. These interfaces serve the OpenMP community as a basis for discussion and prototype implementation for user-defined scheduling.

Siavash Ghiasvand, Florina M. Ciorba

In response to the demand for higher computational power, the number of computing nodes in high performance computers (HPC) increases rapidly. Exascale HPC systems are expected to arrive by 2020. With drastic increase in the number of HPC system components, it is expected to observe a sudden increase in the number of failures which, consequently, poses a threat to the continuous operation of the HPC systems. Detecting failures as early as possible and, ideally, predicting them, is a necessary step to avoid interruptions in HPC systems operation. Anomaly detection is a well-known general purpose approach for failure detection, in computing systems. The majority of existing methods are designed for specific architectures, require adjustments on the computing systems hardware and software, need excessive information, or pose a threat to users' and systems' privacy. This work proposes a node failure detection mechanism based on a vicinity-based statistical anomaly detection approach using passively collected and anonymized system log entries. Application of the proposed approach on system logs collected over 8 months indicates an anomaly detection precision between 62% to 81%.

Siavash Ghiasvand, Florina M. Ciorba

System logs are a valuable source of information for the analysis and understanding of systems behavior for the purpose of improving their performance. Such logs contain various types of information, including sensitive information. Information deemed sensitive can either directly be extracted from system log entries by correlation of several log entries, or can be inferred from the combination of the (non-sensitive) information contained within system logs with other logs and/or additional datasets. The analysis of system logs containing sensitive information compromises data privacy. Therefore, various anonymization techniques, such as generalization and suppression have been employed, over the years, by data and computing centers to protect the privacy of their users, their data, and the system as a whole. Privacy-preserving data resulting from anonymization via generalization and suppression may lead to significantly decreased data usefulness, thus, hindering the intended analysis for understanding the system behavior. Maintaining a balance between data usefulness and privacy preservation, therefore, remains an open and important challenge. Irreversible encoding of system logs using collision-resistant hashing algorithms, such as SHAKE-128, is a novel approach previously introduced by the authors to mitigate data privacy concerns. The present work describes a study of the applicability of the encoding approach from earlier work on the system logs of a production high performance computing system. Moreover, a metric is introduced to assess the data usefulness of the anonymized system logs to detect and identify the failures encountered in the system.

Siavash Ghiasvand, Florina M. Ciorba

System logs constitute valuable information for analysis and diagnosis of system behavior. The size of parallel computing systems and the number of their components steadily increase. The volume of generated logs by the system is in proportion to this increase. Hence, long-term collection and storage of system logs is challenging. The analysis of system logs requires advanced text processing techniques. For very large volumes of logs, the analysis is highly time-consuming and requires a high level of expertise. For many parallel computing centers, outsourcing the analysis of system logs to third parties is the only affordable option. The existence of sensitive data within system log entries obstructs, however, the transmission of system logs to third parties. Moreover, the analytical tools for processing system logs and the solutions provided by such tools are highly system specific. Achieving a more general solution is only possible through the access and analysis system of logs of multiple computing systems. The privacy concerns impede, however, the sharing of system logs across institutions as well as in the public domain. This work proposes a new method for the anonymization of the information within system logs that employs de-identification and encoding to provide sharable system logs, with the highest possible data quality and of reduced size. The results presented in this work indicate that apart from eliminating the sensitive data within system logs and converting them into shareable data, the proposed anonymization method provides 25% performance improvement in post-processing of the anonymized system logs, and more than 50% reduction in their required storage space.