1 Trace-Based Characteristics of Grid Workflows Alexandru Iosup and Dick Epema PDS Group Delft University of Technology The Netherlands Simon Ostermann, Radu Prodan, and Thomas Fahringer DPS Group University of Innsbruck Austria
1
Trace-Based Characteristics of Grid Workflows
Alexandru Iosup and Dick Epema
PDS GroupDelft University of Technology
The Netherlands
Simon Ostermann, Radu Prodan, and Thomas Fahringer
DPS GroupUniversity of Innsbruck
Austria
2
Why are Grid Workflows Interesting?• Grids promise reliable and
easy-to-use computational infrastructure for e-Science
• Full automation from experiment design to final result
• Often, automation = workflows• Jobs comprising inter-related
computing and data-transfer tasks
3
Why are the Characteristics of Grid Workflows Interesting?
• For focusing on the right research problems• What are the interesting characteristics?
Number of nodes? Number of edges? Other characteristics…
• For simulation studies• Optimizing a scheduler for one workload does not make
it useful for another (often quite the contrary)• … optimizing for a workload type is better
• For performance evaluation in real environments• The system tuned to one workload
4
Outline
• Introduction• Method for Grid Workflow Analysis• Austrian Grid Traces• Grid Workflow Characteristics• Conclusion and Future Work
5
Method for Grid Workflow Analysis [1/3]Overview
• Goal: establish the main characteristics of grid workflow such that building a workflow-based grid workload model is greatly facilitated
• Grid workflow characteristics• Workflow-intrinsic• Environment-related
6
Method for Grid Workflow Analysis [2/3]Intrinsic Workflow Characteristics• Size and structure of the workflow
• Number of nodes (N)/edges (E)• Branching Factor = N/E• Work Size = task runtime of a task on a base platform
[SI2k]• Work Size Variability = ratio longest vs. shortest WF task• Sequential execution path• Critical execution path• Graph level (L) = length of critical execution path
• Arrival patterns• Daily patterns: Peak Hours• Weekly patterns: Week-end vs. Work Days
7
Method for Grid Workflow Analysis [3/3]Environment-Related WF Characteristics• Time-related
• Makespan (MS) = time between WF entering and exiting system
• Scheduler-related• Speedup (S) = MS / Sequential Execution Path Size• Normalized Schedule Length (NSL) = MS / Critical Path
Size
• Failure-related• Success rate = % tasks finished correctly, per WF
8
Outline
• Introduction• Method for Grid Workflow Analysis• Austrian Grid Traces• Grid Workflow Characteristics• Conclusion and Future Work
9
The Austrian Grid Traces
• Austrian Grid: 8 sites, ~500 processors• Two non-overlapping long-term traces
from two workflow engines: Askalon DEE, Askalon EE2
• Workflows: mostly testing, but many jobs similar to production workflows
• Production areas: material sciences, astrophysics, weather prediction, engineering, movie rendering
10
Outline
• Introduction• Method for Grid Workflow Analysis• Austrian Grid Traces• Grid Workflow Characteristics• Conclusion and Future Work
11
Intrinsic Workflow Characteristics [1/3]Number of nodes
• 75% WFs have <40 tasks• 95% WFs have < 200 tasks
200 tasks
40 tasks
12
Intrinsic Workflow Characteristics [2/3]Task Work Size
• >80% WFs take <2 minutes on 1000-SI2k machine• >95% WFs take <10 minutes on 1000-SI2k
machine
10 mins
2 mins
14
Classes of Workflows
• Simple classifier (experience from previous work)
• Future: data mining techniques
15
Environment-Related Characteristics
• Workflow class matters: better SU for “easier” classes• Large-and-Flat “easier” than Large-and-Branchy• Large-and-Branchy “easier” than Branchy (o/head)
16
Outline
• Introduction• Method for Grid Workflow Analysis• Austrian Grid Traces• Grid Workflow Characteristics• Conclusion and Future Work
17
Conclusion and Future Work
• Method for the analysis of grid workflows• Intrinsic workflow characteristics• Environment-dependent workflow characteristics• More statistical details than average/std.deviation
(Normal is not the typical distribution in computer science)
• Analysis of two workflow-based traces from Austrian Grid
• Future work• Apply method to more traces• Design workflow-based grid workload model