Multi-SPMD Programming Model with YML and XcalableMP
241
Table 7 Applications’ behaviors and the statuses of error reports
w/ MUST
w/o MUST
Allreduce w/o error
Completed
Completed
Allreduce w/
Completed
Completed
Type conflict error
Error reports
No report
Pingpong w/o error
Completed
Completed
Pingpong w/
Completed
Completed
Type conflict error
Error reports
No report
Allreduce w/
Completed
Completed
Operation conflict error
Error reports
No report
Allreduce w/
Failed
Failed
Buffer size conflict error
Error reports
Simple error reports
0
20
40
60
80
100
NoError-NoMUST NoError-MUST
Error-NoMUST
Error-MUST
0
50
100
150
200
1
2
4
8
16
32
1
2
4
8
16
32
NoError-NoMUST NoError-MUST
Error-NoMUST
Error-MUST
(sec)
(sec)
tasks
tasks
(Allreduce+ 1-sec sleep )x100
(Allreduce+ 0.01-sec sleep )x10000
Fig. 22 Execution time when a workflow includes 1, 2, · · · 32 tasks executing MPI_Allreduce
repeatedly every 1-s (left) and every 0.01-s
0
50
100
150
200
NoError-NoMUST NoError-MUST
Error-NoMUST
Error-MUST
0
50
100
150
200
NoError-NoMUST NoError-MUST
Error-NoMUST
Error-MUST
(sec)
(sec)
tasks
tasks
(Send + 1-sec sleep
Recv+ 1-sec sleep )x50
(Send + 0.01-sec sleep
Recv+ 0.01-sec sleep )x5000
1
2
4
8
16
32
1
2
4
8
16
32
Fig. 23 Execution time when a workflow includes 1, 2, · · · 32 tasks executing MPI_Send/Recv
repeatedly every 1-s (left) and every 0.01-s