Introduction The Design of WaFS The WaFS-aware Scheduler Simulation Study Conclusions and Future Work
On the Benefits of a Workflow-Aware File
System in High-Performance Computing
Systems
Yang Wang Paul Lu
Department of Computing Science University of Alberta, Canada {yangwang, paullu}@cs.ualberta.ca
The 8th International Conference on High Performance Computing in Asia Pacific Region, 2005
Introduction The Design of WaFS The WaFS-aware Scheduler Simulation Study Conclusions and Future Work
Outline
1 Introduction
Workflow-based Workloads in HPC The Problems
Our Contributions
2 The Design of WaFS
The Architecture of WaFS
3 The WaFS-aware Scheduler
Versioned Namespace (VNS) Policy
4 Simulation Study
The Experiment Setup The Results
Introduction
The Design of WaFS The WaFS-aware Scheduler Simulation Study Conclusions and Future Work
Workflow-based Workloads in HPC
The Problems Our Contributions
Workflow-based Workloads in HPC
Characteristics
a set of control/data dependent jobs, multiple instances of the same workflow shape but different input, intermediate, and output files. Example Feature Extraction 2 Function Classifier Localization Classifier Create Summary B C D E F out.B out.C out.D out.E A BLASTout.A out.A Feature Extraction 1
1 a bioinformatics application for proteome classification.
2 input is a large set of DNA/protein sequences.
3 BLAST is to find homologs for a given unknown sequence.
4 extracting features or keyword from the homologs for corresponding classifiers.
Introduction
The Design of WaFS The WaFS-aware Scheduler Simulation Study Conclusions and Future Work
Workflow-based Workloads in HPC
The Problems
Our Contributions
The Problems
1 Filename Conflicts:
1,000 sequences, then 1,000 instances. But, the static filename results in filename conflicts.
Example Out.A A Out.A C D B A Out.A C D Out.A B
Some Known Solutions Serial execution (low concurrency and low performance)
Sub-directories (Sometimes, user’s burden)
Overwrite-Safe Concurrency (OSC) (PDCAT 2005)
Introduction
The Design of WaFS The WaFS-aware Scheduler Simulation Study Conclusions and Future Work
Workflow-based Workloads in HPC The Problems
Our Contributions
Our Contributions
1 The design of a Workflow-aware File System (WaFS):
captures and represents the workflow-specific information (e.g., dataflow information)
2 The design of a WaFS-aware Scheduler:
integration of job scheduler with file system that makes possible a variety of performance optimization and benefits.
3 A simulation study of WaFS and its benefits for job
Introduction
The Design of WaFS
The WaFS-aware Scheduler Simulation Study Conclusions and Future Work
The Architecture of WaFS
The Architecture of WaFS
Data Model
File Work
flow API DepSolver QueryUtil Others
Applications Scheduler Users
Other FS
Traditional File System Run Job User WorkflowSpace WorkflowSpace RunSpace RunSpace JobSpace JobSpace RunSpace RunSpace FileSpace ... UserSpace
The Data Model
Namespace Functionality
UserSpace: User Information, Access Control
WorkflowSpace: Workflow Information (e.g., shape parameters)
RunSpace: Workflow Instance Information (e.g., runtime parameters)
JobSpace: Detail Information of jobs
Introduction
The Design of WaFS
The WaFS-aware Scheduler Simulation Study Conclusions and Future Work
The Architecture of WaFS
The Architecture of WaFS
Data Model
File Work
flow API DepSolver QueryUtil Others
Applications Scheduler Users
Other FS
Traditional File System Run Job User
1 Query Utility: provides remote access services to WaFS.
2 On-line Data Dependency Solver: constructs the data dependencies as the
Introduction The Design of WaFS
The WaFS-aware Scheduler
Simulation Study Conclusions and Future Work
Versioned Namespace (VNS) Policy
Versioned Namespace (VNS) Policy
To enable any existing scheduler to schedule and execute each workflow instance in its own namespace. Example: the execution of WI1, WI2 and WI3 are overlapped and in their own namespaces (NS1, NS2 and NS3). A C B E D F A B C E D F E A F C B D NS2 NS1 NS3 WI2 WI1 WI3
Introduction The Design of WaFS The WaFS-aware Scheduler
Simulation Study
Conclusions and Future Work
The Experiment Setup
The Results
The Experiment Setup
The Workloads
Shape Fork&Join (stage=3, fan-out=32) # of Instances 100 (total 9,600 jobs)
Avg. Interarrival Time exponential distribution (µ≥100 time units) Job service time uniform distribution on [500, 1000] time units
The Policies
Policies Intra-Instance Concurrency Inter-Instance Concurrency BASE Control-flow Impossible in general case
Introduction The Design of WaFS The WaFS-aware Scheduler
Simulation Study
Conclusions and Future Work
The Experiment Setup
The Results
The Results
100 200 400 800 1200 1600 2400 3200 4800 6400 0 1e+05 2e+05 3e+05 4e+05 5e+05 6e+05 7e+05Makespan (Time Units)
BASE (# of files: 128) VNS (# of files: 12800) Fork&Join(3x32) JST[500, 1000]
100 200 400 800 1200 1600 2400 3200 4800 6400 Average Interarrival Time
0 100 200 300 400 500 Avg. DOC.
Introduction The Design of WaFS The WaFS-aware Scheduler
Simulation Study
Conclusions and Future Work
The Experiment Setup
The Results OSC HB DOC Storage Overhead CFG−BASE VNS Sub−dir
Introduction The Design of WaFS The WaFS-aware Scheduler Simulation Study
Conclusions and Future Work
Conclusions and Future Work
1 Introduced the motivation, design, and simulation study of
WaFS.
2 Proposed a Versioned Namespace (VNS) policy to
integrate WaFS and job scheduler for HPC workloads to address the filename conflicts and inefficient
re-computation problems.
3 To address the storage overhead of VNS, we have started
to explore other options, such as the hybrids between dataflow-based policies and VNS.