COMMTEST SP3-375 Point-to-Point Communication Performance

Performance Studies using

COMMTEST

IBM SP3-375 Winterhawk II SWAP Performance

(unordered swap of 8KB message using MPI between two nodes)

(performance measured per processor when all processors in node communicating)

Date/Person:	May 25, 2000 / P. Worley
Platform:	IBM SP3 at Oak Ridge National Laboratory (eagle.ccs.ornl.gov):
	160 4-way Winterhawk II SMP nodes (375MHz POWER3 with 8MB L2 cache)
Environment:	AIX 4.3.3; PSSP 3.1.1
Communication Library:	shared memory MPI within a node, MPI over the switch between nodes, using User Space protocol in both cases
SWAP size:	1024 REAL*8 floating point values each direction
Message size:	Largest - 1024 REAL*8 floating point values
	Smallest - 1 REAL*8 floating point values
Processors:	0 and 4
	1 and 5
	2 and 6
	3 and 7
Latency Definition:	(T₁₀₂₄-T₅₁₂)/512
Model Error Range:	[1,1024]
Results:

unordered simple swap

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 23.33 77.80 22.2%

1 iter. 34.22 79.47 33.2%

10 iter. 31.27 78.72 30.1%

unordered swap using nonblocking send

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 22.56 83.47 23.0%

1 iter. 29.15 84.79 34.7%

10 iter. 31.34 86.63 33.1%

cache inv. w/overlap 21.67 88.83 31.7%

1 iter. w/overlap 33.55 89.31 36.6%

10 iter. w/overlap 31.12 80.74 30.7%

unordered swap using nonblocking receive

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 22.75 84.25 23.4%

1 iter. 30.39 86.56 32.1%

10 iter. 31.47 87.07 33.4%

cache inv. w/overlap 25.73 84.70 27.8%

1 iter. w/overlap 30.56 87.37 32.6%

10 iter. w/overlap 31.56 77.42 29.8%

unordered swap using nonblocking send and receive

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 21.35 91.40 27.0%

1 iter. 35.68 90.81 20.8%

10 iter. 31.39 90.83 34.8%

cache inv. w/overlap 21.43 91.85 30.8%

1 iter. w/overlap 29.47 95.25 34.3%

10 iter. w/overlap 31.29 81.58 31.2%

unordered swap using ready send

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 20.46 159.57 22.1%

1 iter. 27.77 161.03 27.3%

10 iter. 31.43 159.04 30.5%

cache inv. w/overlap 25.57 88.02 27.5%

1 iter. w/overlap 33.12 89.05 28.7%

10 iter. w/overlap 37.69 79.76 18.3%

unordered swap using nonblocking ready send

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 19.95 167.13 26.9%

1 iter. 29.59 170.14 37.3%

10 iter. 31.40 169.66 32.5%

cache inv. w/overlap 25.07 91.45 21.1%

1 iter. w/overlap 30.25 93.78 40.4%

10 iter. w/overlap 38.03 81.78 19.0%

native sendrecv

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 21.72 80.30 24.8%

1 iter. 29.03 81.65 28.9%

10 iter. 31.44 81.97 31.5%

unordered swap using nonblocking sync. send

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 16.02 185.07 35.1%

1 iter. 24.05 189.70 27.8%

10 iter. 27.65 190.05 32.1%

cache inv. w/overlap 17.16 193.16 20.2%

1 iter. w/overlap 23.19 195.65 27.7%

10 iter. w/overlap 27.21 184.04 30.6%

unordered swap using nonblocking receive with sync. send

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 18.66 192.59 24.9%

1 iter. 21.07 196.73 25.5%

10 iter. 26.81 196.53 32.2%

cache inv. w/overlap 16.95 191.20 39.6%

1 iter. w/overlap 22.52 196.64 27.0%

10 iter. w/overlap 27.66 175.71 29.7%

unordered swap using nonblocking sync. send and receive

Data Statistics
bidirectional bandwidth estimated latency model error

(peak MByte/sec) (usec/msg) (max. rel. error)

cache inv. 16.19 204.04 25.5%

1 iter. 20.67 207.06 31.8%

10 iter. 27.10 206.35 34.1%

cache inv. w/overlap 15.89 200.90 30.6%

1 iter. w/overlap 23.22 205.32 29.1%

10 iter. w/overlap 27.13 180.98 30.0%

Protocol Sensitivity Summary for Bidirectional Swap of 8192 Bytes (cache inv./no overlap)
Runtime Statistics
Msg Size min Sec min Sec/Msg max MBytes/Sec (mean-min)/min (median-min)/min (max-min)/min

8 8.2596e-02 8.0660e-05 0.20 0.67 1.00 1.54

16 4.2763e-02 8.3522e-05 0.38 0.64 0.95 1.47

32 2.1730e-02 8.4882e-05 0.75 0.64 0.96 1.46

64 1.1354e-02 8.8701e-05 1.44 0.60 0.94 1.39

128 5.8996e-03 9.2181e-05 2.78 0.66 0.96 1.60

256 3.0689e-03 9.5904e-05 5.34 0.69 0.97 1.66

512 1.6851e-03 1.0532e-04 9.72 0.65 0.85 1.58

1024 1.2120e-03 1.5150e-04 13.52 0.47 0.64 1.08

2048 8.6202e-04 2.1551e-04 19.01 0.34 0.46 0.79

4096 7.0223e-04 3.5112e-04 23.33 0.24 0.30 0.61

8192 8.0074e-04 8.0074e-04 20.46 0.17 0.19 0.30

Five Fastest
Protocols

Msg Size 1st 2nd 3rd 4th 5th

8 0 6 1 2 3

16 0 6 1 2 3

32 0 6 1 2 3

64 0 2 1 6 3

128 0 6 1 2 3

256 0 1 2 6 3

512 2 0 6 1 3

1024 2 0 6 3 1

2048 6 0 2 3 1

4096 0 2 1 6 3

8192 4 5 6 8 1

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 2 5

16 1 4 5

32 1 3 5

64 1 4 5

128 2 5 5

256 1 4 5

512 2 3 5

1024 1 3 5

2048 1 4 5

4096 1 3 5

8192 1 2 6

Protocol Sensitivity Summary for Bidirectional Swap of 8192 Bytes (1 iterations/no overlap)
Runtime Statistics
Msg Size min Sec min Sec/Msg max MBytes/Sec (mean-min)/min (median-min)/min (max-min)/min

8 8.1454e-02 7.9545e-05 0.20 0.70 1.03 1.61

16 4.0767e-02 7.9622e-05 0.40 0.71 1.02 1.61

32 2.0522e-02 8.0165e-05 0.80 0.72 1.04 1.61

64 1.0524e-02 8.2222e-05 1.56 0.69 1.01 1.57

128 5.9999e-03 9.3748e-05 2.73 0.56 0.87 1.34

256 3.0278e-03 9.4618e-05 5.41 0.59 0.83 1.41

512 1.5991e-03 9.9947e-05 10.25 0.64 0.90 1.57

1024 1.0526e-03 1.3157e-04 15.57 0.50 0.33 1.26

2048 6.6894e-04 1.6723e-04 24.49 0.52 0.68 1.15

4096 4.7872e-04 2.3936e-04 34.22 0.45 0.55 0.95

8192 4.5926e-04 4.5926e-04 35.68 0.48 0.64 0.73

Five Fastest
Protocols

Msg Size 1st 2nd 3rd 4th 5th

8 0 6 1 2 3

16 0 6 1 2 3

32 0 6 1 2 3

64 0 6 1 2 3

128 6 3 2 0 1

256 3 6 0 1 2

512 3 2 6 1 0

1024 6 1 3 2 0

2048 1 0 3 2 6

4096 0 3 2 1 6

8192 3 5 4 2 7

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 2 5

16 1 2 5

32 1 1 5

64 1 1 5

128 5 5 5

256 2 5 5

512 2 5 5

1024 1 2 5

2048 1 1 5

4096 1 1 5

8192 1 1 2

Protocol Sensitivity Summary for Bidirectional Swap of 8192 Bytes (10 iterations/no overlap)
Runtime Statistics
Msg Size min Sec min Sec/Msg max MBytes/Sec (mean-min)/min (median-min)/min (max-min)/min

8 8.0984e-02 7.9086e-05 0.20 0.71 1.02 1.61

16 4.0681e-02 7.9454e-05 0.40 0.70 1.02 1.60

32 2.0418e-02 7.9759e-05 0.80 0.71 1.03 1.60

64 1.0456e-02 8.1687e-05 1.57 0.69 1.01 1.55

128 5.8225e-03 9.0976e-05 2.81 0.58 0.90 1.36

256 2.9739e-03 9.2933e-05 5.51 0.60 0.92 1.41

512 1.5685e-03 9.8029e-05 10.45 0.63 0.92 1.56

1024 1.0334e-03 1.2917e-04 15.85 0.54 0.72 1.36

2048 7.0323e-04 1.7581e-04 23.30 0.38 0.50 0.91

4096 5.2067e-04 2.6034e-04 31.47 0.26 0.35 0.65

8192 5.2121e-04 5.2121e-04 31.43 0.13 0.16 0.19

Five Fastest
Protocols

Msg Size 1st 2nd 3rd 4th 5th

8 0 6 1 2 3

16 0 6 1 2 3

32 0 6 1 2 3

64 0 6 1 2 3

128 6 2 0 1 3

256 1 6 2 3 0

512 6 1 2 0 3

1024 1 6 0 2 3

2048 1 6 0 3 2

4096 2 6 3 1 0

8192 4 5 7 1 0

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 2 5

16 1 2 5

32 1 1 5

64 1 1 5

128 3 5 5

256 2 5 5

512 5 5 5

1024 1 5 5

2048 4 5 5

4096 5 5 5

8192 2 2 10

Protocol Sensitivity Summary for Bidirectional Swap of 8192 Bytes (cache inv. with overlap)
Runtime Statistics
Msg Size min Sec min Sec/Msg max MBytes/Sec (mean-min)/min (median-min)/min (max-min)/min

8 8.2963e-02 8.1019e-05 0.20 0.49 0.14 1.50

16 4.2721e-02 8.3439e-05 0.38 0.47 0.12 1.45

32 2.2518e-02 8.7959e-05 0.73 0.42 0.07 1.37

64 1.1487e-02 8.9742e-05 1.43 0.42 0.07 1.39

128 5.9216e-03 9.2525e-05 2.77 0.44 0.03 1.47

256 2.8063e-03 8.7696e-05 5.84 0.61 0.13 1.86

512 1.6635e-03 1.0397e-04 9.85 0.53 0.11 1.60

1024 1.1517e-03 1.4396e-04 14.23 0.39 0.10 1.16

2048 8.0989e-04 2.0247e-04 20.23 0.27 0.06 0.87

4096 6.3681e-04 3.1840e-04 25.73 0.32 0.23 1.00

8192 6.5362e-04 6.5362e-04 25.07 0.43 0.53 0.65

Five Fastest
Protocols

Msg Size 1st 2nd 3rd 4th 5th

8 0 6 2 4 1

16 0 6 2 4 1

32 0 6 2 4 1

64 0 6 2 4 1

128 2 5 0 3 4

256 4 6 0 1 2

512 6 0 2 1 5

1024 4 2 0 6 3

2048 2 6 4 1 5

4096 2 4 5 1 3

8192 5 4 6 3 7

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 7

16 1 2 7

32 1 5 7

64 2 5 7

128 1 7 7

256 1 1 7

512 1 2 7

1024 1 1 7

2048 1 5 7

4096 2 2 6

8192 1 1 2

Protocol Sensitivity Summary for Bidirectional Swap of 8192 Bytes (1 iterations with overlap)
Runtime Statistics
Msg Size min Sec min Sec/Msg max MBytes/Sec (mean-min)/min (median-min)/min (max-min)/min

8 8.2122e-02 8.0197e-05 0.20 0.52 0.17 1.56

16 4.1149e-02 8.0369e-05 0.40 0.52 0.17 1.55

32 2.0803e-02 8.1262e-05 0.79 0.52 0.17 1.54

64 1.0859e-02 8.4834e-05 1.51 0.48 0.11 1.48

128 5.9333e-03 9.2708e-05 2.76 0.39 0.02 1.32

256 2.5235e-03 7.8860e-05 6.49 0.66 0.23 1.80

512 1.6177e-03 1.0110e-04 10.13 0.43 0.01 1.44

1024 1.1088e-03 1.3860e-04 14.78 0.38 0.06 1.21

2048 7.1638e-04 1.7910e-04 22.87 0.33 0.08 1.00

4096 4.8830e-04 2.4415e-04 33.55 0.32 0.14 0.87

8192 4.9468e-04 4.9468e-04 33.12 0.35 0.43 0.49

Five Fastest
Protocols

Msg Size 1st 2nd 3rd 4th 5th

8 0 2 6 4 1

16 0 2 6 4 1

32 0 6 2 1 4

64 0 2 6 4 1

128 6 0 2 1 4

256 5 4 6 3 0

512 0 2 1 3 4

1024 1 3 2 0 5

2048 0 1 6 2 3

4096 1 0 4 2 6

8192 4 5 2 1 3

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 7

16 1 1 7

32 1 1 7

64 1 2 7

128 4 7 7

256 1 1 7

512 3 7 7

1024 3 5 7

2048 1 3 7

4096 1 1 7

8192 1 1 2

Protocol Sensitivity Summary for Bidirectional Swap of 8192 Bytes (10 iterations with overlap)
Runtime Statistics
Msg Size min Sec min Sec/Msg max MBytes/Sec (mean-min)/min (median-min)/min (max-min)/min

8 7.9517e-02 7.7654e-05 0.21 0.43 0.06 1.37

16 3.9877e-02 7.7885e-05 0.41 0.42 0.06 1.37

32 2.0296e-02 7.9283e-05 0.81 0.41 0.05 1.34

64 1.0326e-02 8.0673e-05 1.59 0.41 0.04 1.33

128 5.6240e-03 8.7875e-05 2.91 0.39 0.05 1.26

256 2.9369e-03 9.1778e-05 5.58 0.39 0.03 1.27

512 1.5672e-03 9.7952e-05 10.45 0.41 0.01 1.37

1024 1.0429e-03 1.3036e-04 15.71 0.38 0.03 1.27

2048 6.9796e-04 1.7449e-04 23.47 0.29 0.02 0.94

4096 5.1111e-04 2.5555e-04 32.06 0.21 0.03 0.68

8192 4.3081e-04 4.3081e-04 38.03 0.32 0.40 0.44

Five Fastest
Protocols

Msg Size 1st 2nd 3rd 4th 5th

8 2 4 0 1 3

16 2 4 0 1 5

32 2 4 0 1 3

64 2 4 0 1 5

128 2 4 1 5 3

256 4 2 6 1 3

512 2 4 0 6 3

1024 0 6 4 1 3

2048 4 1 5 0 6

4096 5 4 2 6 3

8192 5 4 8 1 0

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 4 7

16 1 4 7

32 1 6 7

64 1 6 7

128 2 7 7

256 2 7 7

512 4 7 7

1024 3 7 7

2048 2 7 7

4096 2 7 7

8192 2 2 2

DISCUSSION

The SP3 Winterhawk II processing node consists of four processors. In these tests, we examine MPI communication performance between processors in different nodes for communication when all processors in one node are communicating with all processors in the other node. The performance is reported for a single processor pair (logical processors 0 and 4), representing the latency and bandwdith that would typically appear in a highly synchronous communication-intensive operation, such as a collective communication command.