An official website of the United States government
Large-scale computing systems are increasingly using accelerators such as GPUs to enable peta- and exa-scale levels of compute to meet the needs of Machine Learning (ML) and scientific computing applications. Given the widespread and growing use of ML, including in some scientific applications, optimizing these clusters for ML workloads is particularly important. However, recent work has demonstrated that accelerators in these clusters can suffer from performance variability and this variability can lead to resource under-utilization and load imbalance. In this work we focus on how clusters schedulers, which are used to share accelerator-rich clusters across many concurrent ML jobs, can embrace performance variability to mitigate its effects. Our key insight to address this challenge is to characterize which applications are more likely to suffer from performance variability and take that into account while placing jobs on the cluster. We design a novel cluster scheduler, PAL, which uses performance variability measurements and application-specific profiles to improve job performance and resource utilization. PAL also balances performance variability with locality to ensure jobs are spread across as few nodes as possible. Overall, PAL significantly improves GPU-rich cluster scheduling: across traces for six ML workload applications spanning image, language, and vision models with a variety of variability profiles, PAL improves geomean job completion time by 42%, cluster utilization by 28%, and makespan by 47% over existing state-of-the-art schedulers.
more »
« less
The Effect of System Utilization on Application Performance Variability
Application performance variability caused by network contention is a major issue on dragonfly based systems. This work-in-progress study makes two contributions. First, we analyze real workload logs and conduct application experiments on the production system Theta at Argonne to evaluate application performance variability. We find a strong correlation between system utilization and performance variability where a high system utilization (e.g., above 95%) can cause up to 21% degradation in application performance. Next, driven by this key finding, we investigate a scheduling policy to mitigate workload interference by leveraging the fact that production systems often exhibit diurnal utilization behavior and not all users are in a hurry for job completion. Preliminary results show that this scheduling design is capable of improving system productivity (measured by scheduling makespan) as well as improving user-level scheduling metrics such as user wait time and job slowdown.
more »
« less
- PAR ID:
- 10097525
- Date Published:
- Journal Name:
- ACM Digital Library
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Despite extensive investigation of job scheduling in data-intensive computation frameworks, less consideration has been given to optimizing job partitioning for resource utilization and efficient processing. Instead, partitioning and job sizing are a form of dark art, typically left to developer intuition and trial-and-error style experimentation. In this work, we propose that just as job scheduling and resource allocation are out-sourced to a trusted mechanism external to the workload, so too should be the responsibility for partitioning data as a determinant for task size. Job partitioning essentially involves determining the partition sizes to match the resource allocation at the finest granularity. This is a complex, multi-dimensional problem that is highly application specific: resource allocation, computational runtime, shuffle and reduce communication requirements, and task startup overheads all have strong influence on the most effective task size for efficient processing. Depending on the partition size, the job completion time can differ by as much as 10 times! Fortunately, we observe a general trend underlying the tradeoff between full resource utilization and system overhead across different settings. The optimal job partition size balances these two conflicting forces. Given this trend, we design Libra to automate job partitioning as a framework extension. We integrate Libra with Spark and evaluate its performance on EC2. Compared to state-of-the-art techniques, Libra can reduce the individual job execution time by 25% to 70%.more » « less
-
FaaS (Function-as-a-Service) workloads feature unique patterns. Serverless functions are ephemeral, highly concurrent, and bursty, with an execution duration ranging from a few milliseconds to a few seconds. The workload behaviors pose new challenges to kernel scheduling. Linux CFS (Completely Fair Scheduler) is workload-oblivious and optimizes long-term fairness via proportional sharing. CFS neglects the short-term demands of CPU time from short-lived serverless functions, severely impacting the performance of short functions. Preemptive shortest job first—shortest remaining process time (SRPT)—prioritizes shorter functions in order to satisfy their short-term demands of CPU time and, therefore, serves as a best-case baseline for optimizing the turnaround time of short functions. A significant downside of approximating SRPT, however, is that longer functions might be starved. In this paper, we propose a novel application-aware kernel scheduler, ALPS (Adaptive Learning, Priority Scheduler), based on two key insights. First, approximating SRPT can largely benefit short functions but may inevitably penalize long functions. Second, CFS provides necessary infrastructure support to implement user-defined priority scheduling. To this end, we design ALPS to have a novel, decoupled scheduler frontend and backend architecture, which unifies approximate SRPT and proportional-share scheduling. ALPS’ frontend sits in the user space and approximates SRPT-inspired priority scheduling by adaptively learning from an SRPT simulation on a recent past workload. ALPS’ backend uses eBPF functions hooked to CFS to carry out the continuously learned policies sent from the frontend to inform scheduling decisions in the kernel. This design adds workload intelligence to workload-oblivious OS scheduling while retaining the desirable properties of OS schedulers. We evaluate ALPS extensively using two production FaaS workloads (Huawei and Azure), and results show that ALPS achieves a reduction of 57.2% in average function execution duration compared to CFS.more » « less
-
High performance computing (HPC) is undergoing significant changes. The emerging HPC applications comprise both compute- and data-intensive applications. To meet the intense I/O demand from emerging data-intensive applications, burst buffers are deployed in production systems. Existing HPC schedulers are mainly CPU-centric. The extreme heterogeneity of hardware devices, combined with workload changes, forces the schedulers to consider multiple resources (e.g., burst buffers) beyond CPUs, in decision making. In this study, we present a multi-resource scheduling scheme named BBSched that schedules user jobs based on not only their CPU requirements, but also other schedulable resources such as burst buffer. BBSched formulates the scheduling problem into a multi-objective optimization (MOO) problem and rapidly solves the problem using a multi-objective genetic algorithm. The multiple solutions generated by BBSched enables system managers to explore potential tradeoffs among various resources, and therefore obtains better utilization of all the resources. The trace-driven simulations with real system workloads demonstrate that BBSched improves scheduling performance by up to 41% compared to existing methods, indicating that explicitly optimizing multiple resources beyond CPUs is essential for HPC scheduling.more » « less
-
Dragonfly class of networks are considered as promising interconnects for next-generation supercomputers. While Dragonfly+ networks offer more path diversity than the original Dragonfly design, they are still prone to performance variability due to their hierarchical architecture and resource sharing design. Event-driven network simulators are indispensable tools for navigating complex system design. In this study, we quantitatively evaluate a variety of application communication interactions on a 3,456-node Dragonfly+ system by using the CODES toolkit. This study looks at the impact of communication interference from a user’s perspective. Specifically, for a given application submitted by a user, we examine how this application will behave with the existing workload running in the system under different job placement policies. Our simulation study considers hundreds of experiment configurations including four target applications with representative communication patterns under a variety of network traffic conditions. Our study shows that intra-job interference can cause severe performance degradation for communication-intensive applications. Inter-job interference can generally be reduced for applications with one-toone or one-to-many communication patterns through job isolation. Application with one-to-all communication pattern is resilient to network interference.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
