COMMTEST IBM p690 Point-to-Point Communication Performance

Performance Studies using

COMMTEST

IBM p690 SWAP Performance

(unordered swap of 8KB message using MPI within a node)

(performance measured per processor when all processors in an 8 processor LPAR node communicating with neighboring processor)

Date/Person:	April 4, 2002 / P. Worley
Platform:	IBM p690 at Oak Ridge National Laboratory (cheetah.ccs.ornl.gov):
	p690 32-way Turbo SMP node (1.3 GHz POWER4)
Environment:	AIX 5.1L
Communication Library:	MPI
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:	i and i+1, i=0,2,4,...,30
Latency Definition:	(T₁₀₂₄-T₅₁₂)/512
Model Error Range:	[1,1024]
Results:

unordered swap using nonblocking send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 217.03 13.84 33.9%

1 iter. 421.25 13.92 71.6%

10 iter. 531.72 13.96 90.6%

cache inv. w/overlap 213.00 14.33 29.2%

1 iter. w/overlap 393.47 14.33 62.2%

10 iter. w/overlap 540.35 13.04 86.0%

unordered swap using nonblocking receive

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 246.90 10.91 22.2%

1 iter. 419.39 11.03 56.4%

10 iter. 582.04 11.31 80.4%

cache inv. w/overlap 253.31 10.66 33.0%

1 iter. w/overlap 628.22 11.12 85.3%

10 iter. w/overlap 672.98 9.72 79.9%

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. 243.33 14.15 42.0%

1 iter. 420.39 14.12 72.4%

10 iter. 488.96 14.29 85.3%

cache inv. w/overlap 236.67 14.77 42.7%

1 iter. w/overlap 390.84 14.39 68.7%

10 iter. w/overlap 557.68 12.51 85.1%

unordered swap using ready send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 243.33 18.94 28.1%

1 iter. 420.68 18.80 48.3%

10 iter. 675.35 19.04 78.5%

cache inv. w/overlap 311.80 12.35 23.5%

1 iter. w/overlap 642.34 11.84 46.4%

10 iter. w/overlap 1007.87 9.96 61.3%

unordered swap using nonblocking ready send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 211.94 23.31 30.2%

1 iter. 455.79 23.29 64.8%

10 iter. 571.24 23.36 81.4%

cache inv. w/overlap 314.67 14.53 27.9%

1 iter. w/overlap 544.92 14.35 55.7%

10 iter. w/overlap 900.68 12.48 68.6%

native sendrecv

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 294.25 9.58 34.4%

1 iter. 463.00 9.37 53.0%

10 iter. 668.26 9.04 73.7%

unordered swap using nonblocking sync. send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 210.41 26.24 33.7%

1 iter. 356.17 26.46 57.5%

10 iter. 488.26 25.90 77.2%

cache inv. w/overlap 218.14 26.86 35.8%

1 iter. w/overlap 377.16 27.04 62.2%

10 iter. w/overlap 487.58 25.82 76.8%

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. 238.60 25.30 36.8%

1 iter. 414.30 25.01 63.2%

10 iter. 511.83 25.01 78.1%

cache inv. w/overlap 265.34 25.30 41.0%

1 iter. w/overlap 390.84 25.22 60.2%

10 iter. w/overlap 539.16 25.41 83.6%

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. 236.63 28.59 41.3%

1 iter. 353.31 27.63 59.6%

10 iter. 445.33 27.41 74.5%

cache inv. w/overlap 233.75 27.17 46.9%

1 iter. w/overlap 370.79 26.85 60.8%

10 iter. w/overlap 491.60 26.08 78.3%

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 9.9468e-03 9.7136e-06 1.65 0.96 0.95 1.90

16 5.0404e-03 9.8445e-06 3.25 0.94 0.92 1.81

32 2.5858e-03 1.0101e-05 6.34 0.91 0.92 1.79

64 1.2610e-03 9.8518e-06 12.99 0.96 1.00 1.87

128 6.6816e-04 1.0440e-05 24.52 0.88 0.82 1.72

256 3.7711e-04 1.1785e-05 43.45 0.73 0.69 1.51

512 2.2128e-04 1.3830e-05 74.04 0.63 0.72 1.31

1024 1.3449e-04 1.6812e-05 121.82 0.53 0.53 1.13

2048 8.4800e-05 2.1200e-05 193.21 0.45 0.55 0.96

4096 5.5680e-05 2.7840e-05 294.25 0.50 0.57 0.82

8192 6.5733e-05 6.5733e-05 249.25 0.07 0.08 0.13

Five Fastest
Protocols

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

8 6 2 1 3 4

16 6 2 1 3 4

32 6 2 1 3 4

64 6 2 1 3 4

128 6 2 1 3 4

256 6 2 1 3 4

512 2 6 1 3 4

1024 6 2 1 3 4

2048 6 2 1 3 4

4096 6 3 2 1 8

8192 8 2 3 6 1

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 2

16 1 1 2

32 1 1 2

64 1 1 2

128 1 1 2

256 1 1 3

512 1 2 4

1024 1 2 4

2048 1 1 4

4096 1 1 2

8192 1 2 9

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 9.5011e-03 9.2784e-06 1.72 1.03 1.03 1.98

16 4.7031e-03 9.1857e-06 3.48 1.05 1.06 2.00

32 2.4069e-03 9.4019e-06 6.81 1.01 1.03 1.94

64 1.2272e-03 9.5876e-06 13.35 0.98 1.00 1.88

128 6.0081e-04 9.3877e-06 27.27 1.03 1.08 1.92

256 3.0652e-04 9.5787e-06 53.45 1.02 1.06 1.96

512 1.7187e-04 1.0742e-05 95.33 0.84 0.86 1.66

1024 8.8293e-05 1.1037e-05 185.56 0.89 0.91 1.70

2048 5.9373e-05 1.4843e-05 275.95 0.53 0.61 1.14

4096 3.5387e-05 1.7693e-05 463.00 0.47 0.65 0.94

8192 4.0120e-05 4.0120e-05 408.37 0.13 0.17 0.19

Five Fastest
Protocols

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

8 6 2 1 3 4

16 6 2 1 3 4

32 6 2 1 3 4

64 6 2 1 3 4

128 6 2 1 3 4

256 6 2 3 1 4

512 6 2 1 3 4

1024 6 2 1 3 4

2048 2 6 1 3 4

4096 6 1 3 2 4

8192 5 2 9 8 1

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 2

16 1 1 2

32 1 1 2

64 1 1 2

128 1 1 1

256 1 1 1

512 1 1 2

1024 1 1 1

2048 1 2 4

4096 1 1 4

8192 1 1 9

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 9.3117e-03 9.0935e-06 1.76 1.07 1.09 2.02

16 4.6843e-03 9.1491e-06 3.50 1.06 1.08 2.01

32 2.3112e-03 9.0280e-06 7.09 1.08 1.12 2.04

64 1.1705e-03 9.1446e-06 14.00 1.06 1.08 2.01

128 5.8071e-04 9.0735e-06 28.21 1.09 1.13 2.04

256 2.9699e-04 9.2810e-06 55.17 1.05 1.08 1.99

512 1.4628e-04 9.1427e-06 112.00 1.09 1.07 2.08

1024 8.2656e-05 1.0332e-05 198.22 0.91 0.96 1.77

2048 4.5951e-05 1.1488e-05 356.56 0.76 0.75 1.60

4096 2.4517e-05 1.2259e-05 668.26 0.78 0.82 1.60

8192 2.4393e-05 2.4393e-05 671.66 0.31 0.30 0.55

Five Fastest
Protocols

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

8 6 2 1 3 4

16 6 2 1 3 4

32 6 2 1 3 4

64 6 2 1 3 4

128 6 2 1 3 4

256 6 2 1 3 4

512 6 2 1 3 4

1024 6 2 1 3 4

2048 6 2 1 3 4

4096 6 2 1 3 4

8192 4 5 6 2 8

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 2

16 1 1 2

32 1 1 2

64 1 1 2

128 1 1 1

256 1 1 2

512 1 1 1

1024 1 1 2

2048 1 2 3

4096 1 1 2

8192 1 1 3

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 9.9277e-03 9.6951e-06 1.65 0.79 0.51 1.79

16 5.1029e-03 9.9667e-06 3.21 0.75 0.47 1.70

32 2.5796e-03 1.0077e-05 6.35 0.74 0.45 1.71

64 1.3041e-03 1.0188e-05 12.56 0.72 0.44 1.66

128 6.7924e-04 1.0613e-05 24.12 0.70 0.42 1.63

256 3.7899e-04 1.1843e-05 43.23 0.57 0.30 1.40

512 2.2867e-04 1.4292e-05 71.65 0.47 0.18 1.26

1024 1.3448e-04 1.6810e-05 121.83 0.39 0.21 1.01

2048 8.9347e-05 2.2337e-05 183.38 0.28 0.16 0.80

4096 6.4680e-05 3.2340e-05 253.31 0.18 0.09 0.50

8192 4.9893e-05 4.9893e-05 328.38 0.25 0.22 0.43

Five Fastest
Protocols

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

8 6 2 4 1 5

16 6 2 4 1 3

32 6 2 4 1 5

64 6 2 4 1 3

128 6 2 4 1 3

256 6 2 4 5 3

512 2 6 4 1 5

1024 4 2 6 5 1

2048 2 4 5 6 1

4096 2 4 3 6 5

8192 4 5 8 2 3

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 2

16 1 1 2

32 1 1 2

64 1 1 2

128 1 1 3

256 1 1 3

512 1 2 6

1024 1 3 5

2048 1 2 6

4096 1 2 6

8192 1 1 5

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 9.4513e-03 9.2298e-06 1.73 0.84 0.56 1.90

16 4.8341e-03 9.4415e-06 3.39 0.79 0.53 1.82

32 2.3184e-03 9.0561e-06 7.07 0.88 0.59 1.95

64 1.1440e-03 8.9375e-06 14.32 0.93 0.63 2.02

128 5.9771e-04 9.3392e-06 27.41 0.87 0.58 1.92

256 3.3935e-04 1.0605e-05 48.28 0.65 0.41 1.58

512 1.7593e-04 1.0996e-05 93.13 0.67 0.40 1.64

1024 1.0376e-04 1.2970e-05 157.90 0.48 0.26 1.27

2048 5.7280e-05 1.4320e-05 286.03 0.44 0.23 1.06

4096 2.6080e-05 1.3040e-05 628.22 0.77 0.61 1.51

8192 3.2813e-05 3.2813e-05 499.31 0.30 0.34 0.53

Five Fastest
Protocols

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

8 6 2 4 1 3

16 6 2 4 1 3

32 6 2 4 1 3

64 6 2 4 1 3

128 6 2 4 1 5

256 6 2 4 1 3

512 6 4 2 3 1

1024 6 2 4 1 3

2048 6 1 4 2 3

4096 2 6 5 4 3

8192 4 5 2 9 6

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 3

16 1 1 3

32 1 1 2

64 1 1 1

128 1 1 1

256 1 1 4

512 1 1 3

1024 1 2 4

2048 1 1 6

4096 1 1 1

8192 1 1 3

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 9.1707e-03 8.9557e-06 1.79 0.78 0.39 1.92

16 4.5579e-03 8.9021e-06 3.59 0.79 0.39 1.94

32 2.2618e-03 8.8350e-06 7.24 0.81 0.41 1.96

64 1.1485e-03 8.9728e-06 14.27 0.78 0.39 1.91

128 5.7819e-04 9.0342e-06 28.34 0.77 0.36 1.92

256 2.9218e-04 9.1306e-06 56.08 0.77 0.38 1.91

512 1.5480e-04 9.6753e-06 105.84 0.69 0.33 1.76

1024 7.7744e-05 9.7180e-06 210.74 0.72 0.38 1.79

2048 4.3673e-05 1.0918e-05 375.15 0.60 0.28 1.55

4096 2.4345e-05 1.2173e-05 672.98 0.56 0.25 1.42

8192 1.3344e-05 1.3344e-05 1227.82 1.16 1.36 1.54

Five Fastest
Protocols

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

8 6 2 4 3 5

16 6 2 4 3 5

32 6 2 4 3 5

64 6 2 4 3 5

128 6 2 4 3 5

256 6 2 4 5 3

512 6 4 2 3 5

1024 6 4 2 3 1

2048 4 2 6 1 5

4096 2 6 4 3 1

8192 4 5 2 8 6

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 3

16 1 1 3

32 1 1 3

64 1 1 3

128 1 1 3

256 1 1 3

512 1 1 3

1024 1 1 3

2048 1 2 3

4096 1 1 5

8192 1 1 1

DISCUSSION

The p690 processing node consists of 32 processors. To improve connectivity to the network, it can be divided into a number of "logical" smaller nodes, or LPARs. These experiments look at the performance within an 8 processor LPAR, corresponding to a single Multichip Module (MCM).
In these tests, we examine MPI communication performance between neighboring processors in an LPAR node (sharing an L2 cache) when all processors in the node are communicating. The performance is reported for the processor pair with the maximum execution time, representing the worst case latency and bandwidth that would appear in a highly synchronous communication-intensive operation, such as a collective communication command.