Coordinating Workflow Allocation

1
Coordinating Workflow Allocation Execution in
Mobile Environments
Rohan Sen, Gregory Hackmann, Mart Haitjema,
Gruia-Catalin Roman, and Christopher Gill
Mobile Computing LaboratoryDepartment of
Computer Science and Engineering
8th June COORDINATION 2007 Paphos, Cyprus
2
Motivating Example
3
Motivating Example
4
Salient Characteristics of Target Domain

• A structured collaborative activity
• Can be split into discrete smaller tasks
• Each task has requirements and spatiotemporal
  constraints
• Multiple entities involved
• Combination of human skills and software service
• Each entity has a set of capabilities
• The entities are physically mobile
• Mobile Ad hoc Network
• No fixed resources
• Volatile network topology
• Decoupled computing

5
Key Challenges

• Hosts are not always reachable
• Have to be pro-active in allocating tasks
• May not be possible to exchange notifications
  among hosts
• Lack of fixed, centralized execution manager
• Management of workflow execution needs to be
  decentralized
• Data exchange via a peer-to-peer type model
• Diminutive hosts
• Minimalist design philosophy

6
Solution Overview
7
The Knowledge Base

• A local repository on each host
• Contains information about other hosts in the
  MANET
• Location, velocity, motion plan, capabilities
• Populated using a gossiping protocol
• Knowledge is used to determine whether
• Assigning a host to a task causes conflicts with
  other tasks
• Two hosts can communicate during a time interval

8
The Planner

• Responsible for allocating a workflow to a set of
  hosts
• One host in the MANET assumes the role of a
  planner
• Has the complete workflow specification
• Executes a centralized allocation algorithm
• Uses knowledge to make allocation decisions
• Complete knowledge due to initial co-location
• Assume hosts do not report incorrect knowledge
• The allocation problem is NP-Complete
• We use heuristics to reduce time to compute
  allocations

9
Allocation Algorithm

• Step 1 Compute spatiotemporal constraints
• Use information in the knowledge base
• Show a constraint for a host between two tasks as
• Constraints indicate
• Potential conflicts when allocating tasks
• Inability to communicate notifications

10
Allocation Algorithm

• Step 2 Build constraint tables for each task
• Choose hosts that meet the requirements of the
  task
• List these hosts along with their constraints

11
Allocation Algorithm

• Step 3 Sorting the tables
• Sort tables in ascending order of number of rows
• Harder to allocate tasks are allocated first
• Sort rows within tables in ascending order of
  number
• of conflicts
• Hosts with fewer conflicts are preferred

12
Allocation Algorithm

• Step 4 - Perform allocations of tasks
• Use conflict closures to minimize backtracking

?
?
?
?
?
?
Allocate T2 to A
Try to allocate T4
Allocate T4 to D
?
?
Allocate T3 to D
Allocate T3 to B
Allocate T1 to A
3
4
2
13
Solution Overview
14
The Workflow Manager

• Runs on every participating host
• Distributed management of workflow execution
• Aided by Coordination Across Space and Time(CAST)
• Responsible for
• Collecting inputs from the preset of a task
• Perform the task
• Distribute outputs to postset of a task

15
The Role of CAST

• CAST provides basic communication support for
  workflow components
• Three Linda-like operations
• out, in, rd
• Three variants of each operation
• host, loc, reg
• Roman, G.-C., Handorean, R., and Sen, R., Tuple
  Space Coordination Across Space and
• Time,Proceedings of the 8th International
  Conference on Coordination Models and
• Languages (Coordination 2006), Ciancarini, P.,
  and Wiklicky, H., (editors), Lecture
• Notes in Computer Science 4038, Springer,
  Bologna, Italy, June 2006, pp. 266-280.

16
CAST Support for Workflow Operations

• Registering interest in notifications from preset
• Multiple in_host operations
• Propagating results to the postset
• Multiple out_loc operations
• Distributing task allocations to hosts
• out_host for targeted allocations
• out_reg for blanket allocations
• Registering interest in workflow allocations
• rd_host on entity doing the planning
• rd_reg across entire area

17
Related Work

• Workflows have been around for a long time
• Automating repetitive tasks
• Business processes
• Commercial workflow languages and systems
• Business Process Execution Language (BPEL)
• ActiveBPEL, Oracle Workflow, etc.
• Workflow XML (WfXML)
• Systems from academia
• Yet Another Workflow Language (YAWL)
• WORKPAD

18
Experimental Setup

• Randomly generated workflows
• Each data point an average of 50 workflows
• Naïve algorithm used as baseline
• Allocates tasks in workflow order
• Tune-able parameters
• Number of total requirements (r)
• Probability of an agent having a constraint
  between actions (po)
• Number of actions in the plan (a)
• Number of hosts available (g)

19
Algorithm Performance
20
Conclusion

• Allocation algorithm
• Heuristic approach to task allocation
• Use of task closures to minimize backtracking
• Faster than naïve allocation approaches
• Execution environment
• Based on CAST coordination model
• Designed for MANETs
• Future points to consider
• Distributed allocation schemes
• Failures and error handling
• Tradeoff between motion knowledge and performance

21
Questions

• Rohan Sen
• rohan.sen_at_wustl.edu
• www.cse.wustl.edu/ras6
• Mobile Computing Laboratory
• www.cse.wustl.edu/mobilab
• Related Papers
• www.cse.wustl.edu/mobilab/Publications