QoS-CONSTRAINED LIST SCHEDULING HEURISTICS FOR PARALLEL APPLICATIONS ON GRIDS R. BARAGLIA, R.FERRINI, N.TONELLOTTO ISTI, CNR, Pisa, Italy L.RICCI DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF PISA R. YAHYAPOUR INSTITUTE FOR ROBOTICS RESEARCH,

1
QoS-CONSTRAINED LIST SCHEDULING HEURISTICS FOR
PARALLEL APPLICATIONS ON GRIDSR. BARAGLIA,
R.FERRINI, N.TONELLOTTO ISTI, CNR, Pisa,
ItalyL.RICCIDEPARTMENT OF COMPUTER
SCIENCEUNIVERSITY OF PISA R. YAHYAPOUR
INSTITUTE FOR ROBOTICS RESEARCH, UNIVERSITY OF
DORTMUND
2
QLSE DESIGN CHOICES

• QoS-constrained List Scheduling hEuristics for
  Parallel Applications on Grids
• a launch time algorithm to map parallel
  applications on Wide Area Grids
• Basic assumption the user specifies a set of
  Quality of Service
  Requirements
• Computational power
• Communication bandwidth
• Algorithm goals
• To satisfy user QoS requirements
• Fast allocation of tasks to minimize the aging
  effect
• A List Scheduling based solution

3
APPLICATION MODEL

• The application is modelled by a Task Interaction
  Graph (TIG) where
• each node corresponds to a task of the
  application and is associated with the Minimal
  Computational Request (MCR)
• each edge is associated with the Minimal
  Bandwidth Request (MBR)

4
GRID MODEL

• A set of LANs connected through an unreliable
  network
• Each LAN is characterized by
• the number of hosts and the Computational Power
  (CP) of each host
• the internal Communication Bandwidth (CB)
• Two LANs are directly connected in the graph if
  the communication bandwidth
  between them is ? a predefined threshold
  (es1Mb/s)

5
QLSE GENERAL STRUCTURE

• QSLE main goal
• map highly communicating tasks on the same LAN or
  onto a set of LAN connected by high bandwidth
  links
• Overall strategy
• cluster the grid graph so that the LAN belonging
  to each cluster are characterized by high
  communication bandwith
• try to map the application tasks to the hosts of
  the same cluster by a list scheduling approach.
  The mapping must satisfy the QoS specified by the
  user
• if no solution is found, try a further
  clustering characterized by a lower communication
  bandwidth

6
QLSE LIST SCHEDULING

• List Scheduling
• is applied to map the application tasks to the
  LAN of a cluster
• requires an ordered list of the application tasks
  and of LANs belonging to the same cluster
• All application tasks are ordered according to
• the MCR of the Task
• the topology of the TIG
• The LAN available within a cluster are ordered
  according to
• the computational power of the LANs hosts
• the bandwidth of the links between LANs directly
  connected .

7
TASK ORDERING

• 1) Assignment of a priority to a task Ti takes
  into account
• the MCR (Minimal Computational Requirement) of Ti
• the sum of the MBRs (Minimal Bandwith
  Requirement) of the Ti.
• a percentage of sum of the MCRs of the tasks
  interacting with Ti
• priorityi MCRi ?aij?E (MBRij????T MCRj)
• 2) Re-structuring of the TIG into a task
  hierarchical graph THG rooted at
• the highest priority task
• 3) Ordering of the tasks within the same level of
  the THG according to
• the number of communicating tasks
• the value of their priority

8
TASK HIERARCHICAL GRAPH
Tasks 11 and 12 are in the same level because
they have a parent in the previous level
Application TIG
Task 2 and 5 are brothers because they interact
Task Hierarchical Graph

• Cycle Management Tasks belonging to a cycle are
  put in the same cluster

9
GRID CLUSTERING

• Consider the quartiles of the bandwidth
  distribution in decreasing order
• Cluster the grid according to each quartiles
  value, and try to map the tasks on the
  hosts of the same cluster
• If no solution is found, consider the next
  quartile

10
LAN ORDERING

• LAN belonging to the same cluster are ordered
  w.r.t. the priority of the LAN
• The priority value of a LAN Li is computed as
  the sum of
• the computational power of the hosts in Li
• the sum of the bandwidth of the links between Li
  and the directly connected LANs.
• a percentage of sum of the computational power of
  the LANs connected to Li

11
LAN SUITABILITY

• A LAN L is suitable to host a task T iff
• at least a host of L has a computational power
  than the MCR of T
• the sum of the MBRs of the TIG edges between T
  and the communicating tasks already allocated on
  L is ? than the LAN bandwidth
• the sum of the MBRs of the TIG edges between T
  and the communicating tasks already allocated on
  another LAN L' is ??than the link bandwidth
  between L and L
• 1) The computational power of the host where T
  is mapped
• 2) the internal bandwidth of L
• 3) the bandwidth between L and L'
• are decreased according to the corresponding
  values of the TIG

12
QSLE THE ALGORITHM

• Compute the priority of each task
• Build the hierarchical structure of the TIG
  eliminating cyclic paths
• Build the Task Allocation List (TAL)
• Compute the quartile of the grid graphs
• For each quartile
• cluster the grid graph
• rank the cluster by summing the priorities of the
  LANs inside the cluster and build the cluster
  allocation list (CAL)
• for each cluster in CAL
• order the LAN in the clusters according to their
  priority
• select the first task T from TAL and the first
  LAN, from the LAN list, which is suitable for T
• if such a LAN exists, allocate the task on the
  LAN, then consider next task, else consider the
  next cluster
• if no allocation has been found examine next
  quartile

13
AVERAGE PERCENTAGE OF FAILURES

• QLSE has been evaluated through a set of
  simulations
• Greedy Scheduling
• tasks are ordered w.r.t MCR and LAN are ordered
  w.r.t aggregate computational power,the host of
  a LAN ordered w.r.t computational power
• best fit heuristics

14
AVERAGE LAN HIT RATIOS

• LAN hit ratio
• ratio between the sum of the TIG's MBRs of
  communicating tasks
• allocated on the same LAN and the TIG's MBR sum
• measures the percentage of communications
  allocated on the
• same LAN

15
AVERAGE TASK-MACHINE COMPUTATIONAL RATIO

• Task-Machine Computational Ratio
• ratio between the computational power of the
  achine where a
• task is mapped and the MCR of the task
• measueres how powerful machines are exploited to
  run the tasks
• of the TIG

16
CONCLUSIONS

• QLSE a mapping heuristics based on both
  application computational and
• communication requirements (QoS)
• Experimental results demonstrates that QLSE is
  able to carry out a valid solution in almost 100
  of the simulated test cases
• Future work
• A deeper evaluation of QLSE
• An evaluation of QLSE through real applications
  and Grid testbeds