EFFECTIVE LOADBALANCING VIA MIGRATION AND REPLICATION IN SPATIAL GRIDS

1
EFFECTIVE LOAD-BALANCING VIA MIGRATION AND
REPLICATION IN SPATIAL GRIDS

• ANIRBAN MONDAL
• KAZUO GODA
• MASARU KITSUREGAWA
• INSTITUTE OF INDUSTRIAL SCIENCE
• UNIVERSITY OF TOKYO, JAPAN
• anirban,kgoda,kitsure_at_tkl.iis.u-tokyo.ac.jp

2
PRESENTATION OUTLINE

• INTRODUCTION
• RELATED WORK
• SYSTEM OVERVIEW
• MIGRATION AND REPLICATION
• LOAD-BALANCING
• PERFORMANCE STUDY
• CONCLUSION AND FUTURE WORK

3
INTRODUCTION

• Prevalence of spatial applications
• Resource management, development planning,
  emergency planning, scientific research, GIS,
  CAD,VLSI
• Unprecedented growth of available spatial data at
  geographically distributed locations ? the need
  for efficient networking
• Emergence of GRID computing and powerful networks
  
• Motivates the design of a SPATIAL GRID .

4
CHALLENGES

• Scale
• Heterogeneity
• Dynamism
• Cross-domain administrative issues
• Efficient search and load-balancing mechanisms
• We focus on load-balancing.
• Load-balancing in GRIDs is much more complicated
  than in traditional environments.

5
LOAD-BALANCING

• Some nodes become hot
• Skewed Workloads
• Dynamic access patterns
• These hot nodes become bottlenecks
• Increased waiting times ? High response times
• MAIN CONTRIBUTIONS
• Viewing a spatial GRID as comprising several
  clusters
• Each cluster is a LAN
• Proposal of an inter-cluster load-balancing
  algorithm which uses migration/replication of
  data.
• Presentation of a scalable technique for dynamic
  data placement.

6
RELATED WORK

• Ongoing GRID projects
• Earth Systems Grid (ESG)
• NASA Information Power Grid (IPG)
• Grid Physics Network (GriPhyN)
• European DataGrid.
• Binding of execution and storage sites together
  into I/O communities Thain01
• Data-movement system (Kangaroo)
• Load-balancing
• STATIC (BUBBA, tile technique)
• DYNAMIC (Disk cooling)
• Job (Process) MIGRATION in CONDOR
• Spatial indexes R-tree Guttman84

7
SYSTEM OVERVIEW

• Viewing the GRID as a set of clusters
• Distance between two clusters
• Communication time between cluster leaders
• Neighbours
• Definition of Load
• Number of disk I/Os in a certain time interval
• Normalize w.r.t CPU power
• Cluster leaders
• Coordinate cluster activities
• Maintain meta-information
• Data stored at its own cluster its neighbours
• Hotspot detection via access statistics
• Use only recent statistics

8
DATA MOVEMENT IN GRIDs

• MIGRATION REPLICATION
• Unlike replication, migration implies deletion of
  hot data at the source node.
• Which option is better Migration or Replication
• Load-balancing
• Data Availability
• Disk space usage
• Periodic cleanup
• REPLICA CONSISTENCY ??
• Decisions concerning migration/replication should
  be taken during run-time.

9
DATA MOVEMENT (Cont.)

• Impact of heterogeneity on data movement
• Administrative policies (e.g., security)
• Data management techniques (Indexing, hotspot
  detection, etc)
• CPU
• Disk space
• Moving data entails movement of indexes.
• To address variations in indexing schemes, we
  extract data from the index at a node and rebuild
  the index at the destination node.
• Each node has two indexes
• Index for its own data
• Index for moved data

10
DATA MOVEMENT (Cont.)

• Impact of variations in disk space on data
  movement
• Pushing non-hot data to large capacity peers
• Large-sized data migration
• Small-sized data replication
• Replicating small-sized hot data at small
  capacity peers
• Large-sized hot data migration to large capacity
  peers if peers are available, otherwise
  replication.
• Deletion of infrequently accessed replicas

11
INTER-CLUSTER LOAD-BALANCING

• Periodic exchange of load info between neighbours
• Leader L considers itself to be overloaded if its
  load exceeds that of its neighbours by 10.
• L determines its hot regions and informs its
  neighbours about disk space requirement of hot
  regions.
• Number of hot regions depends upon load
  imbalance.
• Neighbours with enough disk space reply to L with
  their load status and disk space information.
• These leaders are sorted (asc) in List1 based on
  their loads.
• L assigns hot regions to members of List 1 in a
  round-robin manner.
• The hottest region is moved to first member of
  List1, the second hottest region is moved to
  second member of List1 and so on.

12
PERFORMANCE STUDY

• 16 SUN workstations, each of which is a 143 MHz
  Sun UltraSparc I processor (256 MB RAM) running
  Solaris 2.5.1 operating system.
• These are connected by relatively high speed
  switch (200 Mbyte/s), the APnet.
• Each cluster is modeled by a workstation node.
• We simulated a transfer rate of 1 Mbit/second
  among the clusters.
• We implemented an R-tree on each of the clusters
  to organize the data allocated to each cluster.
• A real dataset (Greece Roads)
• Each cluster had more than 200000 data
  rectangles.
• Zipf distribution was used to model workload
  skews.
• We investigated only migration in this proposal.

13
PERFORMANCE OF OUR PROPOSED SCHEME
14
SNAPSHOT OF LOAD-BALANCING FOR ZIPF FACTOR OF 0.1
15
VARIATIONS IN WORKLOAD SKEW
16
SNAPSHOT OF LOAD DISTRIBUTION FOR ZIPF FACTOR OF
0.5
17
SUMMARY

• Huge amounts of available spatial data worldwide
  coupled with the emergence of GRID technologies
  and powerful networks motivate the design of a
  spatial GRID.
• For performance reasons, effective load-balancing
  is necessary in such a spatial GRID.
• We view a GRID as a set of clusters.
• Proposal of a dynamic inter-cluster
  load-balancing strategy via migration/replication
  in GRIDs

18
FUTURE SCOPE OF WORK

• FAIRNESS IN LOAD-BALANCING
• GRANULARITY OF DATA MOVEMENT
• DETAILED PERFORMANCE STUDY
• REPLICATION
• DIFFERENT WORKLOAD TYPES
• SCALABILITY
• INTEGRATION INTO EXISTING GRIDs