Source-linked AI summary
Dynamic scheduling of virtual machines running hpc workloads in scientific grids
Omer Khalid, Ivo Maljevic, Richard Anthony, Miltos Petridis, Kevin Parrot, Markus Schulz
TL;DR
The paper addresses uncertain virtualization overhead for deadline-constrained HPC jobs in scientific Grids. It develops a scheduler that monitors deadlines and failure rates through simulation, dynamically deciding whether jobs should continue. The reported outcome is increased job throughput and resource utilization by avoiding premature termination.
Problem
Virtualization overhead varies with workload type, making it difficult to schedule mixed HPC workloads with tight deadlines and competing requirements.
Method
The study develops a virtual-machine scheduling technique that uses simulation, deadline prediction, and real-time failure-rate measurements to guide job decisions.
Results
The scheduler increases overall job throughput and resource utilization by reducing premature termination of executing jobs.
Takeaways & Limitations
Dynamic scheduling can optimize virtualized Grid jobs' deadline success while remaining transparent to users.
Takeaways & Limitations
The control loop can become unstable, requiring threshold values to remain within a specified range.
Abstract
from arXiv · showhide
The primary motivation for uptake of virtualization has been resource isolation, capacity management and resource customization allowing resource providers to consolidate their resources in virtual machines. Various approaches have been taken to integrate virtualization in to scientific Grids especially in the arena of High Performance Computing (HPC) to run grid jobs in virtual machines, thus enabling better provisioning of the underlying resources and customization of the execution environment on runtime. Despite the gains, virtualization layer also incur a performance penalty and its not very well understood that how such an overhead will impact the performance of systems where jobs are scheduled with tight deadlines. Since this overhead varies the types of workload whether they are memory intensive, CPU intensive or network I/O bound, and could lead to unpredictable deadline estimation for the running jobs in the system. In our study, we have attempted to tackle this problem by developing an intelligent scheduling technique for virtual machines which monitors the workload types and deadlines, and calculate the system over head in real time to maximize number of jobs finishing within their agreed deadlines.
1. INTRODUCTION
The paper examines how virtualization overhead affects deadline-constrained HPC jobs in scientific Grids and investigates scheduling for ATLAS grid workloads. It aims to improve job success transparently while preserving utilization efficiency.
- Virtualization overhead may prevent HPC jobs with tight deadlines from meeting their obligations when Grid worker nodes are fully virtualized.
- ATLAS grid infrastructure is investigated by simulating HPC jobs on virtual machines under tight completion deadlines.
- Scientific Grids must serve diverse communities with competing demands, making virtualization overhead and job-duration extensions difficult to manage.
- Scientific clouds may kill jobs that exceed allocated time limits, reducing utilization efficiency.
- The study proposes transparent virtual-machine scheduling intended to optimize job success rates without exposing VM execution to users.
2. MOTIVATION AND BACKGROUND
The paper motivates dynamic, workload-aware scheduling for virtualized scientific-Grid resources and describes its simulation-based scheduling model. The model learns from missed deadlines and adjusts job acceptance using workload and failure-rate information.
- Mixed workloads with competing machine-level requirements remain insufficiently addressed compared with cluster-level utility and cloud resource management.
- Dynamic scheduling must adapt to changing CPU- and memory-intensive workloads and diverse Quality of Service requirements.
- Virtualization provides customizable execution environments, isolation, migration, and fine-grained resource control for Grid workloads.
- The PanDA framework was modified to deploy Grid jobs in virtual machines, with pilot and runjob components managing VM startup, status, execution, and shutdown.
- 2.3 Simulation Model: A simulator was developed because running thousands of ATLAS jobs, some lasting up to 24 hours, in virtual machines was impractical.
- 2.3 Simulation Model: The scheduler learns about missed deadlines and predicts whether jobs will meet deadlines before taking measures to improve success chances.
- 2.3 Simulation Model: The x factor represents the ratio for a job projected to miss its current deadline and is used in acceptance decisions against threshold X.
- 2.3 Simulation Model: The threshold method accepts jobs when x_i < X and rejects them otherwise.
3. CASE STUDY
The case study simulates HPC workloads under physical, static-virtualization, and dynamically managed virtual-machine configurations. Adaptive and statistical thresholding improve deadline-oriented job success and failure-rate behavior compared with less intelligent virtualization management.
- Configurations: The simulator evaluates five configurations: physical baseline, static virtualization, dynamic virtualization, adaptive-threshold management, and statistical-threshold management.The study uses these configurations to compare workload execution under different virtualization-overhead management strategies.
- Evaluation metrics: Figure 2 reports overall job success rate and deadline miss rate for each tested algorithm.The figure uses solid and broken lines for the two measures.
- Performance results: 0.78 to 0.84, a 7.7% increase in job success rate, is achieved when adaptive and statistical dynamic configurations are activated.The result is reported for the transition from alg_3 to the enhanced configurations.
- Threshold behavior: Alg_4 raises X as failure rate rises and lowers it to 0.1, while alg_5 keeps X within [0.1, 0.3) and its failure rate below 0.79.The two dynamic strategies differ in how their thresholds respond and stabilize during simulation.
4. RELATED WORK
The related work situates this study among Xen scheduling, datacenter VM-management, and Grid execution efforts. Its distinguishing focus is deadline-aware HPC job scheduling rather than interactive latency or datacenter consolidation.
- Xen scheduling: Prior Xen scheduler work includes approaches focused on interactive workloads and comparisons among different Xen schedulers.The cited work provides scheduling context for virtual-machine resource management.
- Scientific Grid context: The present work extends existing LHC Grid infrastructure experience to enable on-demand virtual-machine-based job execution on the Grid.It builds on deployment frameworks and infrastructure knowledge from the scientific Grid community.
- Scope comparison: Unlike VSched’s real-time and user-interactive scheduling focus, this work targets batch HPC workloads with little user interactivity.The related-work comparison distinguishes the workload and scheduling objective addressed here.
- Datacenter management: Datacenter studies dynamically migrate virtual workloads to manage hotspots, temperature, operational cost, and consolidation under peak load.These approaches consider migration and resource placement across datacenter servers.
- Scope boundary: Live migration to rescue jobs before deadline expiration remains outside this study’s scope and is identified as future research.The authors connect migration research to deadline-aware scheduling but do not include it in the current work.
5. CONCLUSION
The paper presents a dynamic, adaptive VM scheduler for Grid HPC workloads that estimates virtualization overhead in real time and makes deadline-aware continuation or termination decisions. It reports improved job success and system throughput, with integration into batch or cluster schedulers.
- The technique dynamically schedules virtual machines for HPC workloads while accounting for virtualization performance impact and deadline obligations.The scheduler uses real-time overhead estimation to guide scheduling decisions.
- Real-time failure-rate measurement helps decide whether a running job should continue or be terminated based on its chances of meeting the deadline.
- Prematurely terminating jobs that would otherwise never complete increases overall job success rate and system throughput.
- The VM scheduler can integrate with batch or cluster schedulers, which select resource compositions while node-level scheduling optimizes job success.