COMMTEST IBM p690 Point-to-Point Communication Performance

Performance Studies using

COMMTEST

IBM p690 SWAP Performance

(unordered swap of 2MB message using MPI within a node)

(performance measured per processor when all processors in node communicating with a "distant" 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:	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:	i and i+16, i=0,...,15
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. 406.07 2.28 40.0%

1 iter. 434.42 2.10 43.2%

10 iter. 835.37 -0.05 65.2%

cache inv. w/overlap 403.69 1.57 42.2%

1 iter. w/overlap 441.37 1.87 44.8%

10 iter. w/overlap 839.66 0.09 65.0%

unordered swap using nonblocking receive

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 410.64 0.23 45.0%

1 iter. 441.55 0.82 45.7%

10 iter. 836.75 -0.69 66.7%

cache inv. w/overlap 409.72 0.68 43.3%

1 iter. w/overlap 439.36 0.52 46.0%

10 iter. w/overlap 845.81 -0.69 67.3%

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. 409.73 -4.96 75.3%

1 iter. 435.73 -4.51 75.5%

10 iter. 687.17 -0.93 65.3%

cache inv. w/overlap 407.95 -3.52 68.2%

1 iter. w/overlap 439.54 -3.51 70.7%

10 iter. w/overlap 832.02 0.54 64.3%

unordered swap using ready send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 410.85 17.84 8.3%

1 iter. 436.63 17.58 7.3%

10 iter. 862.87 18.29 47.5%

cache inv. w/overlap 404.52 9.51 5.4%

1 iter. w/overlap 439.34 9.72 5.5%

10 iter. w/overlap 950.31 10.06 52.0%

unordered swap using nonblocking ready send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 405.13 19.31 8.4%

1 iter. 434.19 19.30 8.9%

10 iter. 852.92 19.57 46.4%

cache inv. w/overlap 405.37 11.78 7.6%

1 iter. w/overlap 440.21 11.80 7.3%

10 iter. w/overlap 934.04 12.02 51.3%

native sendrecv

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 409.60 -0.17 44.7%

1 iter. 434.21 -0.44 47.7%

10 iter. 850.39 -2.00 75.8%

unordered swap using nonblocking sync. send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 408.09 23.04 14.2%

1 iter. 437.40 24.11 17.1%

10 iter. 837.71 21.81 45.7%

cache inv. w/overlap 407.25 23.20 15.2%

1 iter. w/overlap 437.21 23.31 14.8%

10 iter. w/overlap 836.98 21.97 45.5%

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. 406.11 21.91 15.4%

1 iter. 438.96 21.10 12.3%

10 iter. 838.96 20.78 45.5%

cache inv. w/overlap 410.75 21.63 13.9%

1 iter. w/overlap 437.30 21.54 15.4%

10 iter. w/overlap 844.23 20.80 45.1%

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. 406.70 21.87 20.7%

1 iter. 443.98 20.39 24.8%

10 iter. 682.85 23.60 30.3%

cache inv. w/overlap 406.87 20.19 21.1%

1 iter. w/overlap 436.98 19.82 21.9%

10 iter. w/overlap 830.79 24.58 45.0%

Protocol Sensitivity Summary for Bidirectional 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 1.8205e-02 1.7778e-05 230.39 0.40 0.48 0.91

4096 1.7852e-02 3.4867e-05 234.95 0.11 0.05 0.32

8192 1.3768e-02 5.3782e-05 304.64 0.09 0.03 0.36

16384 1.1914e-02 9.3075e-05 352.06 0.09 0.02 0.35

32768 1.1164e-02 1.7444e-04 375.70 0.08 0.01 0.34

65536 1.0748e-02 3.3588e-04 390.24 0.05 0.01 0.20

131072 1.0761e-02 6.7254e-04 389.78 0.03 0.04 0.08

262144 1.0620e-02 1.3275e-03 394.94 0.01 0.01 0.02

524288 1.0362e-02 2.5905e-03 404.78 0.01 0.01 0.02

1048576 1.0237e-02 5.1184e-03 409.73 0.01 0.01 0.01

2097152 1.0425e-02 1.0425e-02 402.33 0.01 0.01 0.01

Five Fastest
Protocols

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

2048 6 2 1 3 4

4096 4 6 8 5 2

8192 6 4 2 8 5

16384 6 4 8 2 1

32768 6 8 4 7 1

65536 4 5 8 1 2

131072 4 1 7 5 8

262144 1 4 7 9 5

524288 5 4 9 3 2

1048576 3 7 6 9 1

2097152 2 5 6 3 7

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 1 2 4

4096 1 4 7

8192 1 7 7

16384 3 7 7

32768 6 7 7

65536 5 7 9

131072 4 6 9

262144 4 9 9

524288 6 9 9

1048576 5 9 9

2097152 8 9 9

Protocol Sensitivity Summary for Bidirectional 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 1.7617e-02 1.7204e-05 238.08 0.42 0.50 0.89

4096 1.7348e-02 3.3884e-05 241.77 0.11 0.05 0.32

8192 1.3238e-02 5.1711e-05 316.84 0.10 0.02 0.37

16384 1.1390e-02 8.8985e-05 368.24 0.08 0.01 0.36

32768 1.0463e-02 1.6348e-04 400.87 0.09 0.01 0.37

65536 1.0084e-02 3.1513e-04 415.94 0.05 0.02 0.21

131072 1.0182e-02 6.3636e-04 411.94 0.03 0.04 0.05

262144 9.8034e-03 1.2254e-03 427.84 0.02 0.03 0.04

524288 9.4470e-03 2.3617e-03 443.98 0.03 0.04 0.05

1048576 9.4703e-03 4.7351e-03 442.89 0.02 0.02 0.04

2097152 9.3780e-03 9.3780e-03 447.25 0.02 0.02 0.04

Five Fastest
Protocols

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

2048 6 2 1 3 4

4096 4 6 5 8 2

8192 6 4 8 5 2

16384 8 6 2 4 5

32768 8 4 6 7 1

65536 8 4 5 2 6

131072 7 5 4 1 9

262144 7 4 5 8 3

524288 9 8 3 6 7

1048576 9 8 7 4 3

2097152 6 7 9 5 2

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 1 1 4

4096 1 4 7

8192 1 7 7

16384 4 7 7

32768 2 7 7

65536 2 7 9

131072 4 7 9

262144 2 9 9

524288 1 9 9

1048576 2 9 9

2097152 2 9 9

Protocol Sensitivity Summary for Bidirectional 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 1.1912e-02 1.1633e-05 352.11 0.63 0.79 1.30

4096 1.1994e-02 2.3426e-05 349.70 0.15 0.11 0.28

8192 7.5053e-03 2.9318e-05 558.85 0.11 0.07 0.26

16384 5.5110e-03 4.3055e-05 761.08 0.11 0.05 0.36

32768 4.8609e-03 7.5952e-05 862.87 0.10 0.03 0.38

65536 5.2193e-03 1.6310e-04 803.62 0.05 0.01 0.18

131072 6.1567e-03 3.8479e-04 681.26 0.03 0.01 0.07

262144 6.6362e-03 8.2953e-04 632.03 0.11 0.10 0.17

524288 7.7154e-03 1.9289e-03 543.63 0.11 0.13 0.17

1048576 8.6074e-03 4.3037e-03 487.29 0.05 0.05 0.08

2097152 8.8384e-03 8.8384e-03 474.55 0.04 0.05 0.05

Five Fastest
Protocols

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

2048 6 2 1 3 4

4096 4 5 6 2 8

8192 4 6 5 2 8

16384 4 6 5 2 8

32768 4 5 6 8 7

65536 1 4 7 5 6

131072 2 9 6 3 8

262144 3 9 2 8 6

524288 3 9 8 2 6

1048576 3 9 2 8 6

2097152 9 3 8 5 6

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 1 1 4

4096 1 1 7

8192 1 3 7

16384 1 5 7

32768 1 7 7

65536 4 7 9

131072 5 5 9

262144 2 2 9

524288 2 2 9

1048576 2 2 9

2097152 2 7 9

Protocol Sensitivity Summary for Bidirectional 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 1.8110e-02 1.7685e-05 231.61 0.30 0.16 0.80

4096 1.4452e-02 2.8226e-05 290.23 0.30 0.29 0.58

8192 1.2154e-02 4.7476e-05 345.10 0.20 0.16 0.51

16384 1.1225e-02 8.7694e-05 373.66 0.13 0.07 0.41

32768 1.0696e-02 1.6712e-04 392.15 0.12 0.05 0.38

65536 1.0701e-02 3.3441e-04 391.95 0.06 0.02 0.19

131072 1.0705e-02 6.6906e-04 391.81 0.03 0.05 0.06

262144 1.0552e-02 1.3190e-03 397.50 0.02 0.02 0.04

524288 1.0447e-02 2.6118e-03 401.48 0.01 0.00 0.01

1048576 1.0323e-02 5.1615e-03 406.31 0.01 0.00 0.02

2097152 1.0449e-02 1.0449e-02 401.41 0.01 0.01 0.02

Five Fastest
Protocols

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

2048 6 2 4 1 3

4096 4 5 6 2 8

8192 4 5 6 2 8

16384 4 5 6 8 1

32768 4 5 6 2 7

65536 4 6 5 7 1

131072 4 5 7 1 2

262144 4 5 7 1 9

524288 9 8 4 2 3

1048576 8 7 2 6 4

2097152 8 6 3 4 7

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 1 1 6

4096 1 1 2

8192 1 2 7

16384 1 2 7

32768 1 4 7

65536 3 6 9

131072 2 4 9

262144 3 9 9

524288 8 9 9

1048576 8 9 9

2097152 8 9 9

Protocol Sensitivity Summary for Bidirectional 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 1.7414e-02 1.7006e-05 240.86 0.32 0.17 0.85

4096 1.3765e-02 2.6884e-05 304.72 0.32 0.32 0.61

8192 1.1622e-02 4.5399e-05 360.89 0.21 0.17 0.49

16384 1.0649e-02 8.3195e-05 393.87 0.14 0.08 0.42

32768 1.0108e-02 1.5794e-04 414.95 0.11 0.04 0.39

65536 9.7497e-03 3.0468e-04 430.20 0.09 0.05 0.22

131072 9.9339e-03 6.2087e-04 422.22 0.04 0.05 0.07

262144 9.7801e-03 1.2225e-03 428.86 0.02 0.02 0.05

524288 9.5984e-03 2.3996e-03 436.98 0.01 0.02 0.03

1048576 9.5685e-03 4.7842e-03 438.35 0.01 0.01 0.03

2097152 9.4593e-03 9.4593e-03 443.41 0.01 0.00 0.03

Five Fastest
Protocols

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

2048 6 2 4 1 3

4096 4 5 6 8 2

8192 4 5 6 8 2

16384 4 5 6 8 2

32768 4 5 6 8 2

65536 4 5 7 6 8

131072 4 5 7 1 8

262144 5 4 6 7 9

524288 9 3 8 6 4

1048576 2 8 9 5 3

2097152 7 6 4 3 9

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 1 1 6

4096 1 1 2

8192 1 2 7

16384 2 2 7

32768 1 5 7

65536 1 5 9

131072 3 5 9

262144 2 9 9

524288 2 9 9

1048576 3 9 9

2097152 7 9 9

Protocol Sensitivity Summary for Bidirectional 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 1.1844e-02 1.1567e-05 354.12 0.50 0.28 1.32

4096 8.1026e-03 1.5825e-05 517.65 0.58 0.66 0.84

8192 5.6686e-03 2.2143e-05 739.92 0.38 0.43 0.55

16384 4.6990e-03 3.6711e-05 892.59 0.20 0.23 0.28

32768 4.4136e-03 6.8963e-05 950.31 0.10 0.13 0.14

65536 4.8601e-03 1.5188e-04 863.00 0.06 0.07 0.08

131072 6.0747e-03 3.7967e-04 690.45 0.03 0.00 0.08

262144 6.9136e-03 8.6420e-04 606.68 0.06 0.04 0.12

524288 8.2517e-03 2.0629e-03 508.29 0.05 0.05 0.09

1048576 8.9778e-03 4.4889e-03 467.19 0.02 0.02 0.03

2097152 9.1566e-03 9.1566e-03 458.06 0.01 0.01 0.01

Five Fastest
Protocols

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

2048 6 2 4 1 3

4096 4 5 6 2 8

8192 4 5 6 2 8

16384 4 5 6 2 8

32768 4 5 6 2 8

65536 4 5 1 6 2

131072 8 6 3 9 2

262144 8 2 3 9 6

524288 8 2 3 9 6

1048576 2 8 9 3 6

2097152 6 9 2 8 3

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 1 1 4

4096 1 1 2

8192 1 1 2

16384 1 2 5

32768 1 2 9

65536 1 2 9

131072 5 7 9

262144 2 5 9

524288 2 5 9

1048576 2 9 9

2097152 7 9 9

DISCUSSION

The p690 processing node consists of 32 processors. In these tests, we examine MPI communication performance between processors I and (I+16) in a node (preventing L2 cache haring and snooping) 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.