PSTSWM AlphaSC-667 Point-to-Point Communication Performance

Performance Studies using

PSTSWM

Compaq AlphaServer SC SWAP Performance

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

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

< td> Latency Definition:

Date/Person:	July 17, 2000 / P. Worley
Platform:	Compaq AlphaServer SC at Oak Ridge National Laboratory (falcon.ccs.ornl.gov):
	64 ES40 4-way SMP nodes (667 MHz Alpha 21264a with 8MB L2 cache)
Environment:	SC System v1.0; Digital UNIX V5.0; RMS 2.37
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:	0 and 1
	2 and 3
(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. 65.91 36.08 47.7%

1 iter. 25.96 37.44 96.4%

10 iter. 168.63 36.38 55.2%

unordered swap using nonblocking send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 86.87 38.74 47.6%

1 iter. 115.22 39.02 50.7%

10 iter. 162.41 38.92 53.2%

cache inv. w/overlap 81.67 38.46 45.8%

1 iter. w/overlap 67.37 38.36 64.6%

10 iter. w/overlap 167.18 39.14 73.4%

unordered swap using nonblocking receive

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 86.60 32.69 51.9%

1 iter. 119.94 35.09 53.7%

10 iter. 171.17 33.19 57.3%

cache inv. w/overlap 99.18 22.38 60.3%

1 iter. w/overlap 129.62 22.53 62.7%

10 iter. w/overlap 171.60 25.71 69.8%

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. 89.14 29.89 54.7%

1 iter. 121.72 32.48 55.1%

10 iter. 173.82 34.15 55.3%

cache inv. w/overlap 97.87 23.16 59.4%

1 iter. w/overlap 123.19 22.59 62.7%

10 iter. w/overlap 173.41 24.96 64.0%

unordered swap using ready send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 73.01 63.18 35.5%

1 iter. 102.79 63.25 39.7%

10 iter. 126.73 63.35 49.0%

cache inv. w/overlap 101.01 22.00 61.8%

1 iter. w/overlap 145.51 22.40 64.1%

10 iter. w/overlap 178.01 24.85 64.3%

unordered swap using nonblocking ready send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 76.20 65.98 33.1%

1 iter. 103.57 64.41 40.7%

10 iter. 127.03 65.20 50.6%

cache inv. w/overlap 98.94 22.61 61.7%

1 iter. w/overlap 145.25 22.45 63.2%

10 iter. w/overlap 172.64 27.08 61.6%

native sendrecv

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 88.85 35.08 49.7%

1 iter. 134.74 29.87 59.0%

10 iter. 167.46 32.59 58.1%

unordered swap using nonblocking sync. send

Data Statistics
bidirectional bandwidth estimated latency model error

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

cache inv. 79.38 36.67 49.7%

1 iter. 125.07 36.88 53.3%

10 iter. 160.31 37.29 54.1%

cache inv. w/overlap 80.79 45.62 56.8%

1 iter. w/overlap 124.12 47.89 59.0%

10 iter. w/overlap 147.50 39.31 66.2%

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. 86.60 37.52 50.7%

1 iter. 131.49 37.79 52.8%

10 iter. 162.57 37.48 54.5%

cache inv. w/overlap 93.84 37.19 51.3%

1 iter. w/overlap 136.53 37.54 52.9%

10 iter. w/overlap 162.86 37.71 55.8%

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. 84.37 37.85 49.4%

1 iter. 126.81 37.89 52.8%

10 iter. 163.61 37.84 54.3%

cache inv. w/overlap 95.15 31.67 54.7%

1 iter. w/overlap 136.31 31.76 57.2%

10 iter. w/overlap 159.72 33.31 57.2%

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 3.0590e-02 2.9873e-05 0.54 0.38 0.26 1.19

16 1.5289e-02 2.9861e-05 1.07 0.37 0.26 1.16

32 7.4424e-03 2.9072e-05 2.20 0.40 0.28 1.22

64 4.4650e-03 3.4883e-05 3.67 0.34 0.21 1.01

128 2.1496e-03 3.3588e-05 7.62 0.45 0.37 1.18

256 1.2678e-03 3.9619e-05 12.92 0.35 0.34 0.93

512 1.4602e-03 9.1263e-05 11.22 0.10 0.08 0.31

1024 7.5400e-04 9.4250e-05 21.73 0.13 0.10 0.36

2048 4.0820e-04 1.0205e-04 40.14 0.14 0.15 0.36

4096 2.4520e-04 1.2260e-04 66.82 0.15 0.15 0.35

8192 1.8380e-04 1.8380e-04 89.14 0.10 0.06 0.35

Five Fastest
Protocols

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

8 3 2 6 0 7

16 3 6 2 0 7

32 3 2 6 0 8

64 2 6 0 7 8

128 3 2 6 0 7

256 3 2 6 0 1

512 3 2 6 1 7

1024 3 2 6 9 1

2048 2 3 1 6 8

4096 2 1 6 3 8

8192 3 6 1 2 8

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 1 5

16 1 1 4

32 1 1 4

64 1 1 6

128 1 1 3

256 2 3 3

512 1 3 8

1024 1 3 8

2048 1 4 8

4096 1 3 8

8192 2 5 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 3.2852e-02 3.2082e-05 0.50 0.29 0.17 1.02

16 1.6184e-02 3.1609e-05 1.01 0.30 0.19 1.06

32 8.0852e-03 3.1583e-05 2.03 0.30 0.18 1.04

64 4.1388e-03 3.2334e-05 3.96 0.37 0.27 1.12

128 2.1150e-03 3.3047e-05 7.75 0.42 0.35 1.16

256 1.2592e-03 3.9350e-05 13.01 0.33 0.32 0.84

512 1.4568e-03 9.1050e-05 11.25 0.11 0.07 0.29

1024 7.4460e-04 9.3075e-05 22.00 0.28 0.05 1.93

2048 3.8200e-04 9.5500e-05 42.89 5.61 0.09 55.09

4096 2.0660e-04 1.0330e-04 79.30 0.32 0.11 2.06

8192 1.2160e-04 1.2160e-04 134.74 10.71 0.12 105.90

Five Fastest
Protocols

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

8 6 3 2 0 7

16 2 3 6 0 8

32 2 6 3 0 7

64 2 6 3 7 8

128 3 6 2 0 8

256 3 2 6 0 7

512 3 6 2 7 8

1024 6 3 2 9 8

2048 6 3 2 8 9

4096 6 8 3 9 2

8192 6 8 9 7 3

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 2 3 8

16 1 2 8

32 1 2 8

64 1 3 3

128 1 2 3

256 1 2 3

512 2 3 8

1024 2 4 7

2048 1 5 7

4096 1 2 7

8192 1 2 7

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 3.3670e-02 3.2881e-05 0.49 0.26 0.14 0.97

16 1.6222e-02 3.1684e-05 1.01 0.30 0.18 1.04

32 8.1584e-03 3.1869e-05 2.01 0.28 0.17 1.02

64 4.3744e-03 3.4175e-05 3.75 0.32 0.21 1.03

128 2.2144e-03 3.4600e-05 7.40 0.37 0.29 1.07

256 1.2639e-03 3.9497e-05 12.96 0.33 0.31 0.87

512 1.4338e-03 8.9613e-05 11.43 0.10 0.07 0.29

1024 7.2894e-04 9.1117e-05 22.48 0.08 0.06 0.28

2048 3.5870e-04 8.9675e-05 45.68 0.11 0.08 0.31

4096 1.8070e-04 9.0350e-05 90.67 0.10 0.07 0.34

8192 9.4260e-05 9.4260e-05 173.82 0.11 0.07 0.37

Five Fastest
Protocols

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

8 6 3 2 0 7

16 3 2 6 0 7

32 3 6 2 0 7

64 3 2 6 7 0

128 2 6 3 0 8

256 3 2 6 0 7

512 3 2 6 0 7

1024 3 0 6 2 1

2048 3 2 6 0 9

4096 2 3 6 1 0

8192 3 2 0 6 9

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 3 3 8

16 1 3 8

32 2 3 8

64 1 1 7

128 1 2 4

256 1 3 4

512 1 3 8

1024 2 5 8

2048 1 3 8

4096 1 4 8

8192 1 4 8

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 2.3026e-02 2.2486e-05 0.71 0.38 0.41 1.08

16 1.1589e-02 2.2635e-05 1.41 0.39 0.48 1.11

32 5.6772e-03 2.2177e-05 2.89 0.39 0.44 1.17

64 3.4030e-03 2.6586e-05 4.81 0.35 0.36 1.28

128 1.9412e-03 3.0331e-05 8.44 0.31 0.31 1.15

256 1.1846e-03 3.7019e-05 13.83 0.25 0.28 0.99

512 1.3186e-03 8.2413e-05 12.43 0.15 0.14 0.64

1024 6.9000e-04 8.6250e-05 23.74 0.14 0.11 0.48

2048 3.8300e-04 9.5750e-05 42.78 0.11 0.12 0.31

4096 2.3040e-04 1.1520e-04 71.11 0.12 0.10 0.35

8192 1.6220e-04 1.6220e-04 101.01 0.12 0.08 0.46

Five Fastest
Protocols

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

8 4 5 2 3 6

16 3 5 4 2 9

32 5 3 2 4 6

64 3 5 2 4 6

128 5 4 3 2 6

256 3 5 2 4 6

512 2 3 4 5 6

1024 5 3 4 2 9

2048 4 2 5 3 9

4096 2 4 3 5 9

8192 4 2 5 3 9

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 2 4 4

16 2 4 4

32 1 4 4

64 2 4 5

128 3 4 5

256 2 4 5

512 2 4 9

1024 1 4 9

2048 2 3 8

4096 3 4 8

8192 1 4 8

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 2.2898e-02 2.2361e-05 0.72 0.40 0.45 1.15

16 1.1403e-02 2.2271e-05 1.44 0.40 0.46 1.17

32 5.6218e-03 2.1960e-05 2.91 0.40 0.45 1.26

64 3.3120e-03 2.5875e-05 4.95 0.39 0.38 1.36

128 1.8932e-03 2.9581e-05 8.65 0.34 0.34 1.24

256 1.1100e-03 3.4688e-05 14.76 0.32 0.34 1.10

512 1.2834e-03 8.0213e-05 12.77 0.19 0.14 0.71

1024 6.6180e-04 8.2725e-05 24.76 1.00 0.13 8.63

2048 3.5620e-04 8.9050e-05 46.00 0.60 0.08 4.89

4096 1.9540e-04 9.7700e-05 83.85 0.21 0.07 1.20

8192 1.1260e-04 1.1260e-04 145.51 1.14 0.12 6.06

Five Fastest
Protocols

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

8 5 4 2 3 9

16 5 4 3 2 9

32 5 4 3 2 6

64 5 2 4 3 6

128 3 5 2 4 6

256 4 2 5 3 6

512 5 2 4 3 6

1024 4 5 3 2 6

2048 5 4 2 3 6

4096 4 5 6 3 2

8192 4 5 8 9 6

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 4 4 4

16 1 4 4

32 1 4 4

64 1 4 4

128 2 4 5

256 1 2 5

512 1 4 8

1024 1 3 8

2048 1 4 8

4096 1 2 8

8192 2 2 8

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 2.5597e-02 2.4997e-05 0.64 0.27 0.32 0.58

16 1.2241e-02 2.3908e-05 1.34 0.32 0.37 0.66

32 6.1226e-03 2.3916e-05 2.68 0.30 0.36 0.66

64 3.4418e-03 2.6889e-05 4.76 0.33 0.36 0.84

128 1.9322e-03 3.0190e-05 8.48 0.31 0.34 0.96

256 1.2029e-03 3.7590e-05 13.62 0.25 0.25 0.92

512 1.3705e-03 8.5658e-05 11.95 0.22 0.11 0.99

1024 6.7744e-04 8.4680e-05 24.19 0.13 0.09 0.64

2048 3.4672e-04 8.6680e-05 47.25 0.12 0.10 0.60

4096 1.7810e-04 8.9050e-05 91.99 0.11 0.09 0.45

8192 9.2040e-05 9.2040e-05 178.01 0.07 0.06 0.21

Five Fastest
Protocols

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

8 4 3 2 5 6

16 5 2 3 4 6

32 3 2 5 4 6

64 2 3 5 4 6

128 3 5 2 4 6

256 4 6 3 5 2

512 3 4 5 6 9

1024 4 3 2 5 6

2048 4 2 5 3 6

4096 0 4 2 5 3

8192 4 3 5 0 2

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 2 4 4

16 1 2 4

32 1 3 4

64 2 3 4

128 1 2 5

256 1 4 5

512 3 4 8

1024 1 4 9

2048 2 4 9

4096 1 3 9

8192 1 5 10

DISCUSSION

The AlphaServer SC processing node consists of four processors. In these tests, we examine MPI communication performance between processors in a node 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.