MPI COLLECTIVE COMMUNICATIONS FOR GRIDS broadcast

1
MPI COLLECTIVE COMMUNICATIONS FOR
GRIDS(broadcast)

• Progress Report
• Rakhi Gupta
• MSc (Engg.) SERC

2
Motivation

• Collective operations
• Topology aware based on hierarchy (MPICH-G)
• MagPIe one WAN link policy
• Indirect link might be faster than the direct
  link

3
Motivation (continued)

• Static techniques - Communication tree does not
  dynamically adapt with changes in network
  characteristics.
• Objective To develop a system that
• Converges to the optimal tree.
• Adaptive

4
Key observation

• Broadcast operation with the same of set of nodes
  and the same root may occur many times, within
  the same application or even across applications.

Per fomance
Per fomance
Topology aware Broadcast
Optimized broadcast
time
time
First run of application
Ith run of application
5
Strategy
Single Step transformation trans (2,4)
6
6
4
4
3
3
Select 2 random numbers (2,4)
2
2
5
5
Single step modified tree
Initial tree
6
Algorithm

• Start with a flat tree t0.
• bad_link_list f.
• Tree_list t0.
• Till performance is not satisfactory,
• tj trans (a,b) ti , (a,b) ? bad_link_list i
  gt 0 i lt size(Tree_list)
• If time(tj) gt time(ti) bad_link_list
  bad_link_list (a,b).
• If time(tj) lt MIN_TIME(Tree_list) threshold
• Remove MAX_TIME(Tree_list) from Tree_list.
• Add tj to Tree_list.

7
Architecture
Service synchronization of experiments and MPI
Application. Experiment finds the bcast
tree. MPI Application uses the tree made by
experiment
IPC through shared memory
Bcast service

• One experiment can execute at one instant
• Many applications may execute simultaneously.

8
Scenario 1
2. j Find_exec_exp()
3. e
Bcast service
4. s
1. p
5. g
Stopping an experiment at j to run broadcast
with root i
9
Scenario 2
2. k Find_recent() 3. Update(i)
4. b
Bcast service
1. f
When broadcast at i finishes, experiment k starts
10
Broadcast Implementation

• If node is Root node
• Note time t1
• Send to direct successors data sub tree for
  that successor root info
• Wait for acknowledgement from all the leaf nodes.
• Note time t2 (all ack obtained)
• Bcast time is t2-t1
• Else
• Wait for data from parent
• If node is not leaf node
• Send to direct successors data sub tree for
  that successor root info
• Else
• Send Acknowledgement to root (null message
  packets)
• For getting proper results experimental
  broadcasts are repeated many times.

11
References

• Thilo Kielmann, Rutger F.H. Hofman, Henri E. Bal,
  Aske Plaat, and Raoul A.F. Bhoedjang MagPIe
  MPI's Collective Communication Operations for
  Clustered Wide Area Systems
• Nicholas T Karonis, Ian Foster, et al Exploiting
  hierarchy in parallel computer networks to
  optimize collective operation performance.