Streamline: A Scheduling Heuristic for Streaming Applications on the Grid

1
Streamline A Scheduling Heuristic for
StreamingApplications on the Grid

• Bikash Agarwalla
• Nova Ahmed, David Hilley
• Kishore Ramachandran
• College of Computing
• Georgia Tech

2
Summary

• Streaming applications are ubiquitous
• Resource management for streaming applications is
  a relatively unexplored territory
• Present a heuristic, called Streamline, for
  scheduling streaming applications in Grid
• Comparison with Optimal, Simulated Annealing, and
  a baseline scheduler, called E-Condor, shows
  favorable results !

3
Motivation
Applications Environment Monitoring,
Distributed Surveillance, Emergency Response,
Habitat Monitoring
4
Motivation
3rd
2nd
Coarse-grain Dataflow Graph
3rd

• Streaming Application Dataflow Graph
• Require efficient use of High Performance
  Computing (HPC) resources
• Comparison with Task-Graph scheduling in
  Multi-Processor Systems list scheduling,
  Job-Shop Scheduling of 70s
• A Mapping Issue, not an Ordering Issue

5
Grid Computing

• Provide access to ambient HPC resources
• Initial focused on scientific and engineering
  applications
• Some recent efforts for Interactive and Streaming
  applications (Interactive Grid HP , GATES
  OSU)
• Resource management infrastructure primarily for
  batch-oriented applications

6
Scheduling Problem

• Input
• (1) Computation and communication requirements of
  various stages of a coarse-grain dataflow graph
• (2) Application-specified constraints
• (3) Current resource (processing and bandwidth)
  availability
• (4) Resource specific constraints
• Output
• Placement of the stages of the pipeline on
  available HPC resources
• Performance criteria
• latency and throughput of the application
• Use existing grid technology tools and standards
• Result a periodic scheduler

7
Scheduling Problem

• Application Model
• Continuous Coarse-grain Dataflow Graph
• Estimates of
• Average CPU cycles required by a stage
• Average data communication between stages
• Assign costs to nodes and edges as in Equation 1
  and 2.
• Estimates can be provided by the application or
  derived by profiling

8
Scheduling Problem

• Available Resources
• Static Information (Infrequently changing
  information)
• Dynamic Information (frequently changing
  information)
• Use Existing Tools Grid Information Service
  (Globus Toolkit), Network Weather Service (UCSB)

Ni
Bi,j
Procj
Nj
9
Scheduling Problem

• Application Specific Constraints
• Co-locate specific stages of the application on
  same node
• Special requirements of a stage (need for a
  graphics co-processor)
• QoS requirements policies (desired throughput,
  latency)
• Resource Specific Constraints
• Site specific policies
• Obtained from the Grid Information Service
  (Globus Toolkit)
• Output
• Mapping stages of dataflow graph to available
  resources
• Performance Criteria
• Maximize total throughput (rate at which data
  items are produced by the exit stages)

10
Solution Streamline
R0
R1
R3
Resource Filtering
R2
S2 R0 R2 R3
Resource Selection
S2 R0

• Belongs to the general class of List Scheduling
  algorithms, but there are differences

11
Solution Streamline

• Stage Prioritization
• Give higher rank to a more needy stage (denoted
  by rank)
• In case of tie, give a stage closer to the entry
  stages a higher priority (blevel)
• For stages with same rank and blevel,
  tie-breaking is done randomly
• Consider the stages in the order of priority

12
Solution Streamline

• Resource Filtering
• Filter out available resources that are not
  permissible by application and resource policies
• Result A candidate set R of r candidate
  resources
• Resource Selection Phase
• For each stage si in order of the priority,
  assign a resource nj with minimum cost A(nj,si)
• Time complexity O(v.r2)

13
Solution Streamline

• Expects to maximize throughput by assigning best
  resource to most needy stage
• Additional policies concerning resources,
  applications, and local schedulers can be
  incorporated in the cost of a particular
  assignment
• Streamline provides
• APIs for users to specify resource requirements
  of each stage and dependencies
• APIs for job submission allowing multiple
  applications to be submitted at the same time
• APIs for querying job status using a
  distinguished name

14
System Architecture
15
Evaluation

• Enhance an existing grid scheduler to deal with
  streaming applications (resulted in a baseline
  stream scheduler called E-Condor)
• Optimal Algorithm for small dataflow graph
• Simulated Annealing for small and large dataflow
  graphs
• Comparison for executing kernels of streaming
  applications
• Scalability studies of Streamline with respect to
  Optimal, Simulated Annealing, and E-Condor

16
E-Condor Architecture
17
Optimal Placement

• Exhaustive search over all possible assignment of
  resources to individual stages
• Cost of an assignment relates to throughput of
  the application
• Run time complexity (rPv)

18
Simulated Annealing

• The state space is all possible assignment of
  candidate resources to the stages
• Use the same cost function as in the Optimal
• Two different versions of SA (SA1, and SA2) with
  different run-time overheads
• SA2 has more run-time overhead than SA1, and
  gives better result.

19
Experimental Parameters

• Algorithms Streamline, E-Condor, Optimal,
  Simulated Annealing (SA1, SA2)
• Application Workload Collage,EdgeD, MotionD,
  FD/FR. Computation and Communication Numbers
  from Basenet04 (Wolenetz et al.) paper.
• Two different 4 stage dataflow graphs for
  micro-measurements and 15 stage dataflow graph
  for scalability study.
• Resource Contention Control Variables Mean and
  Variance in processing and bandwidth availability
  (µp,sp2, µbw,sbw2)
• CPU Contention Modeling Introduced synthetic
  delay in the code depending on the (assumed) load
  on the machine on which it is running
• Network Bandwidth Contention Modeling Inflate
  the data size between the stages in proportion to
  the level of (assumed) network contention
• Intent of experiments Compare quality of node
  selection, measured in terms of throughput
• Experimental platform 17 node cluster (Pentium
  III 550Mhz Xeon processors with 4GB RAM)

20
Micro measurements

• Use 4 nodes of the cluster (1 processor in each
  node)
• Two kernels of distributed surveillance
  application
• Compute-bound
• Communication-bound
• Comparison Metric Average time taken per output
  data item averaged over 100 items

21
Compute-bound Kernel
22
Compute-bound Kernel (CPU Availability Variance)
sp0.65 µbw1 sbw20
23
Communication-bound Kernel
24
Communication-bound Kernel (Bandwidth
Availability Variance)
µp 1 sp2 0 µbw 0.58
25
Results from Micro measurements

• Performance of Streamline is comparable to
  (within 1) Optimal and SA2 algorithm
• Streamline performs significantly better than
  E-Condor especially under non-uniform load
  conditions
• For communication bound kernel, SA2 performance
  is better than SA1

26
Scalability
27
Scalability Results
28
Related Work

• Grid Scheduling Legion, Nimrod-G, Condor, GATES
• Task-graph Scheduling List scheduling and its
  variants HLF (Highest Level First), LP (Longest
  Path), LPT (Longest Processing Time), CP
  (Critical Path)
• Stream processing in database research Aurora,
  TelegraphCQ

29
Conclusion

• Resource management for streaming application is
  new territory.
• We presented Streamline heuristic and E-Condor
  architecture as a baseline grid scheduler for
  streaming application
• Comparison with Optimal, SA, and E-Condor shows
  favorable results
• Streamline outperforms baseline, E-Condor, by an
  order of magnitude for both compute and
  communication bound kernels, particularly under
  non-uniform load conditions
• Streamline performs close to Simulated Annealing
  with very low scheduling overhead (by a factor of
  1000)

30
Thank You ! Contact bikash_at_cc.gatech.edu