COMMTEST SP3-375 Point-to-Point Communication Performance

Performance Studies using

COMMTEST

IBM SP3-375 Winterhawk II SWAP Performance

(ordered swap of 2MB 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:	262144 REAL*8 floating point values each direction
Message size:	Largest - 262144 REAL*8 floating point values
	Smallest - 256 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:

ordered simple swap

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.36 67.13 29.0%

1 iter. 41.00 69.71 28.1%

10 iter. 41.24 70.28 28.0%

ordered swap using nonblocking send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.26 67.20 29.2%

1 iter. 41.11 69.86 28.3%

10 iter. 41.28 70.58 28.1%

cache inv. w/overlap 40.31 71.49 28.0%

1 iter. w/overlap 41.02 70.49 28.4%

10 iter. w/overlap 41.19 71.44 27.9%

ordered swap using nonblocking receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.34 71.73 27.7%

1 iter. 40.98 69.27 28.5%

10 iter. 41.17 70.51 28.0%

cache inv. w/overlap 40.50 70.55 28.6%

1 iter. w/overlap 41.14 70.23 28.6%

10 iter. w/overlap 41.17 71.40 27.7%

ordered swap using nonblocking send and receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.29 70.92 28.4%

1 iter. 40.96 72.02 27.8%

10 iter. 41.25 70.69 28.1%

cache inv. w/overlap 40.38 71.22 28.5%

1 iter. w/overlap 41.14 73.35 27.9%

10 iter. w/overlap 41.22 72.64 28.1%

ordered swap using ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.27 80.36 12.3%

1 iter. 41.01 79.67 9.4%

10 iter. 41.09 79.09 8.6%

cache inv. w/overlap 40.46 71.70 12.6%

1 iter. w/overlap 41.18 70.18 5.6%

10 iter. w/overlap 41.34 71.28 9.5%

ordered swap using nonblocking ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.41 81.34 10.8%

1 iter. 41.15 81.27 7.9%

10 iter. 41.31 79.88 8.7%

cache inv. w/overlap 40.49 74.09 7.2%

1 iter. w/overlap 41.24 71.84 8.7%

10 iter. w/overlap 41.32 71.26 8.6%

synchronous

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.19 76.39 27.5%

1 iter. 41.00 74.98 27.9%

10 iter. 41.29 77.33 27.2%

ordered swap using nonblocking sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.34 196.71 5.4%

1 iter. 40.99 197.05 5.7%

10 iter. 41.24 193.98 5.4%

cache inv. w/overlap 40.22 197.88 4.5%

1 iter. w/overlap 41.11 197.78 4.8%

10 iter. w/overlap 41.19 193.54 5.5%

ordered swap using nonblocking receive with sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.29 197.89 5.4%

1 iter. 41.02 196.72 5.5%

10 iter. 41.23 193.74 5.1%

cache inv. w/overlap 40.32 196.50 5.0%

1 iter. w/overlap 41.04 197.76 5.8%

10 iter. w/overlap 41.19 194.57 5.2%

ordered swap using nonblocking sync. send and receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.52 196.76 4.5%

1 iter. 41.12 196.15 5.0%

10 iter. 41.24 192.99 5.2%

cache inv. w/overlap 40.41 198.21 5.1%

1 iter. w/overlap 41.18 197.39 5.2%

10 iter. w/overlap 41.19 193.69 4.9%

ordered simple swap using sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 40.54 197.39 5.5%

1 iter. 41.19 196.43 5.4%

10 iter. 41.29 190.53 4.9%

Protocol Sensitivity Summary for Unidirectional Swap of 2097152 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

2048 2.4212e-01 2.3645e-04 17.32 0.40 0.07 1.05

4096 1.7086e-01 3.3370e-04 24.55 0.28 0.03 0.73

8192 1.2935e-01 5.0528e-04 32.43 0.42 0.51 0.53

16384 1.1907e-01 9.3023e-04 35.23 0.22 0.26 0.27

32768 1.1846e-01 1.8510e-03 35.41 0.10 0.12 0.13

65536 1.1340e-01 3.5437e-03 36.99 0.05 0.05 0.06

131072 1.0825e-01 6.7654e-03 38.75 0.02 0.03 0.03

262144 1.0567e-01 1.3208e-02 39.69 0.01 0.01 0.02

524288 1.0449e-01 2.6121e-02 40.14 0.01 0.01 0.01

1048576 1.0405e-01 5.2024e-02 40.31 0.00 0.00 0.01

2097152 1.0347e-01 1.0347e-01 40.54 0.01 0.00 0.01

Five Fastest
Protocols

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

2048 0 1 2 3 6

4096 2 0 1 3 4

8192 4 5 10 7 2

16384 5 4 9 10 0

32768 4 5 2 7 0

65536 4 5 1 10 0

131072 4 5 9 1 7

262144 5 4 10 2 1

524288 5 4 0 1 2

1048576 5 10 3 4 1

2097152 10 9 5 0 7

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 3 4 7

4096 1 7 7

8192 1 2 2

16384 2 2 2

32768 2 2 11

65536 2 2 11

131072 2 11 11

262144 2 11 11

524288 11 11 11

1048576 11 11 11

2097152 10 11 11

Protocol Sensitivity Summary for Unidirectional Swap of 2097152 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

2048 2.4270e-01 2.3701e-04 17.28 0.40 0.07 1.04

4096 1.7132e-01 3.3461e-04 24.48 0.27 0.03 0.72

8192 1.3079e-01 5.1090e-04 32.07 0.39 0.47 0.49

16384 1.1490e-01 8.9762e-04 36.51 0.22 0.27 0.28

32768 1.1698e-01 1.8279e-03 35.85 0.10 0.11 0.13

65536 1.1146e-01 3.4832e-03 37.63 0.05 0.06 0.06

131072 1.0670e-01 6.6690e-03 39.31 0.03 0.03 0.03

262144 1.0400e-01 1.3000e-02 40.33 0.01 0.02 0.03

524288 1.0280e-01 2.5700e-02 40.80 0.01 0.01 0.01

1048576 1.0212e-01 5.1060e-02 41.07 0.00 0.01 0.01

2097152 1.0182e-01 1.0182e-01 41.19 0.00 0.00 0.01

Five Fastest
Protocols

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

2048 0 2 1 3 6

4096 0 3 1 2 6

8192 4 5 10 0 1

16384 5 4 10 1 0

32768 5 4 0 10 2

65536 4 5 8 10 0

131072 4 5 8 0 10

262144 5 4 9 0 2

524288 4 5 1 7 2

1048576 5 8 4 7 6

2097152 10 5 9 1 6

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 3 5 7

4096 4 7 7

8192 1 2 2

16384 2 2 2

32768 2 2 11

65536 2 2 11

131072 2 11 11

262144 2 11 11

524288 11 11 11

1048576 11 11 11

2097152 11 11 11

Protocol Sensitivity Summary for Unidirectional Swap of 2097152 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

2048 2.4222e-01 2.3654e-04 17.32 0.39 0.06 1.04

4096 1.7026e-01 3.3253e-04 24.64 0.28 0.04 0.74

8192 1.3245e-01 5.1737e-04 31.67 0.37 0.45 0.46

16384 1.1221e-01 8.7663e-04 37.38 0.23 0.28 0.30

32768 1.1508e-01 1.7981e-03 36.45 0.10 0.12 0.13

65536 1.1059e-01 3.4558e-03 37.93 0.05 0.06 0.07

131072 1.0628e-01 6.6423e-03 39.47 0.02 0.02 0.04

262144 1.0363e-01 1.2954e-02 40.47 0.01 0.01 0.03

524288 1.0207e-01 2.5518e-02 41.09 0.01 0.00 0.02

1048576 1.0153e-01 5.0763e-02 41.31 0.00 0.00 0.01

2097152 1.0159e-01 1.0159e-01 41.29 0.00 0.00 0.00

Five Fastest
Protocols

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

2048 0 2 1 3 6

4096 0 2 1 3 6

8192 4 5 0 2 8

16384 5 4 0 2 10

32768 4 5 10 0 1

65536 4 5 0 10 9

131072 4 5 10 0 8

262144 4 5 0 8 3

524288 4 5 0 2 3

1048576 5 10 1 3 8

2097152 6 10 7 0 9

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 4 4 7

4096 4 7 7

8192 2 2 2

16384 2 2 2

32768 2 2 11

65536 2 2 11

131072 2 11 11

262144 9 11 11

524288 10 11 11

1048576 9 11 11

2097152 11 11 11

Protocol Sensitivity Summary for Unidirectional Swap of 2097152 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

2048 2.4160e-01 2.3594e-04 17.36 0.40 0.02 1.07

4096 1.7128e-01 3.3453e-04 24.49 0.27 0.01 0.72

8192 1.2471e-01 4.8716e-04 33.63 0.47 0.57 0.58

16384 1.1739e-01 9.1708e-04 35.73 0.23 0.28 0.29

32768 1.1778e-01 1.8403e-03 35.61 0.11 0.13 0.13

65536 1.1220e-01 3.5063e-03 37.38 0.06 0.07 0.07

131072 1.0792e-01 6.7450e-03 38.86 0.03 0.03 0.03

262144 1.0537e-01 1.3172e-02 39.80 0.01 0.02 0.02

524288 1.0453e-01 2.6134e-02 40.12 0.00 0.01 0.01

1048576 1.0370e-01 5.1849e-02 40.45 0.00 0.00 0.01

2097152 1.0355e-01 1.0355e-01 40.50 0.00 0.00 0.01

Five Fastest
Protocols

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

2048 0 2 1 4 3

4096 5 1 4 2 0

8192 4 5 0 8 9

16384 4 5 8 7 3

32768 5 4 0 1 10

65536 4 5 7 9 0

131072 5 4 9 7 1

262144 4 5 10 7 3

524288 5 10 4 1 0

1048576 5 0 4 3 8

2097152 2 5 4 0 9

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 1 6 7

4096 5 7 7

8192 1 1 2

16384 1 2 2

32768 2 2 11

65536 2 2 11

131072 2 11 11

262144 2 11 11

524288 11 11 11

1048576 11 11 11

2097152 11 11 11

Protocol Sensitivity Summary for Unidirectional Swap of 2097152 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

2048 2.4168e-01 2.3602e-04 17.35 0.39 0.02 1.06

4096 1.7104e-01 3.3407e-04 24.52 0.27 0.02 0.73

8192 1.2744e-01 4.9782e-04 32.91 0.43 0.52 0.55

16384 1.1311e-01 8.8364e-04 37.08 0.24 0.29 0.30

32768 1.1491e-01 1.7955e-03 36.50 0.11 0.13 0.14

65536 1.1099e-01 3.4684e-03 37.79 0.05 0.06 0.07

131072 1.0719e-01 6.6991e-03 39.13 0.02 0.03 0.03

262144 1.0401e-01 1.3001e-02 40.33 0.01 0.02 0.02

524288 1.0253e-01 2.5633e-02 40.91 0.01 0.01 0.01

1048576 1.0198e-01 5.0992e-02 41.13 0.01 0.01 0.01

2097152 1.0171e-01 1.0171e-01 41.24 0.00 0.00 0.01

Five Fastest
Protocols

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

2048 0 2 1 4 3

4096 0 3 2 1 4

8192 5 4 8 0 9

16384 5 4 0 7 2

32768 4 5 0 8 6

65536 5 4 1 10 8

131072 5 4 8 7 10

262144 4 5 1 3 10

524288 5 4 1 6 7

1048576 4 1 5 2 8

2097152 5 9 4 2 3

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 2 6 7

4096 5 7 7

8192 1 2 2

16384 2 2 2

32768 2 2 11

65536 2 2 11

131072 2 11 11

262144 2 11 11

524288 5 11 11

1048576 10 11 11

2097152 11 11 11

Protocol Sensitivity Summary for Unidirectional Swap of 2097152 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

2048 2.4244e-01 2.3675e-04 17.30 0.38 0.01 1.04

4096 1.7041e-01 3.3283e-04 24.61 0.27 0.01 0.74

8192 1.2807e-01 5.0028e-04 32.75 0.41 0.50 0.52

16384 1.0931e-01 8.5396e-04 38.37 0.26 0.31 0.32

32768 1.1403e-01 1.7818e-03 36.78 0.11 0.13 0.14

65536 1.1009e-01 3.4404e-03 38.10 0.05 0.06 0.06

131072 1.0614e-01 6.6338e-03 39.52 0.02 0.03 0.03

262144 1.0341e-01 1.2926e-02 40.56 0.01 0.01 0.01

524288 1.0232e-01 2.5579e-02 40.99 0.00 0.00 0.01

1048576 1.0147e-01 5.0735e-02 41.34 0.00 0.00 0.01

2097152 1.0155e-01 1.0155e-01 41.30 0.00 0.00 0.00

Five Fastest
Protocols

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

2048 0 2 1 4 5

4096 0 2 1 4 3

8192 4 5 0 10 8

16384 4 5 0 1 10

32768 5 4 10 0 2

65536 4 5 0 1 10

131072 4 5 8 10 7

262144 5 4 3 1 9

524288 5 3 10 7 0

1048576 4 5 6 0 10

2097152 5 4 0 6 10

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 4 6 7

4096 5 7 7

8192 1 2 2

16384 2 2 2

32768 2 2 11

65536 2 2 11

131072 2 11 11

262144 4 11 11

524288 10 11 11

1048576 11 11 11

2097152 11 11 11

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.