PSTSWM AlphaSC-667 Point-to-Point Communication Performance

Performance Studies using

PSTSWM

Compaq AlphaServer SC SWAP Performance

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

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

> 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 4 nodes
	1 and 5
	2 and 6
	3 and 7
(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. 33.64 12.34 25.6%

1 iter. 35.48 12.26 21.1%

10 iter. 36.86 12.19 21.2%

ordered swap using nonblocking send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 32.62 11.80 20.7%

1 iter. 35.84 12.29 21.1%

10 iter. 36.82 12.14 22.2%

cache inv. w/overlap 32.60 12.56 21.6%

1 iter. w/overlap 35.31 12.29 22.6%

10 iter. w/overlap 36.61 12.66 21.9%

ordered swap using nonblocking receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 33.53 12.78 22.6%

1 iter. 35.82 12.37 22.9%

10 iter. 36.77 12.54 22.0%

cache inv. w/overlap 34.06 12.52 21.9%

1 iter. w/overlap 36.26 12.22 25.1%

10 iter. w/overlap 37.01 15.56 17.4%

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. 33.27 12.71 22.4%

1 iter. 35.85 12.69 22.0%

10 iter. 36.74 12.63 22.0%

cache inv. w/overlap 34.02 12.68 22.7%

1 iter. w/overlap 35.69 12.69 22.8%

10 iter. w/overlap 37.07 15.32 21.5%

ordered swap using ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 31.41 28.05 15.5%

1 iter. 34.87 28.37 14.4%

10 iter. 35.70 28.08 14.8%

cache inv. w/overlap 33.42 12.40 22.7%

1 iter. w/overlap 35.87 12.87 21.6%

10 iter. w/overlap 37.24 15.33 18.8%

ordered swap using nonblocking ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 31.11 28.95 14.4%

1 iter. 33.91 28.94 17.4%

10 iter. 35.65 28.72 15.2%

cache inv. w/overlap 33.13 12.78 21.7%

1 iter. w/overlap 36.34 12.58 24.1%

10 iter. w/overlap 36.97 15.09 22.0%

synchronous

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 32.51 27.08 13.6%

1 iter. 34.54 27.15 15.7%

10 iter. 36.19 26.33 13.2%

ordered swap using nonblocking sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 32.81 19.48 15.2%

1 iter. 35.91 19.52 13.4%

10 iter. 36.80 19.27 14.2%

cache inv. w/overlap 32.14 19.13 12.4%

1 iter. w/overlap 36.04 19.78 11.5%

10 iter. w/overlap 36.64 20.33 14.8%

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. 33.21 19.35 13.6%

1 iter. 35.60 19.03 13.7%

10 iter. 36.91 19.17 11.3%

cache inv. w/overlap 33.66 18.96 12.3%

1 iter. w/overlap 35.73 19.31 12.4%

10 iter. w/overlap 37.19 20.03 10.3%

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. 32.61 19.41 11.5%

1 iter. 35.80 19.47 12.5%

10 iter. 36.67 19.06 12.2%

cache inv. w/overlap 33.03 19.04 11.6%

1 iter. w/overlap 35.98 19.52 9.2%

10 iter. w/overlap 36.93 20.59 11.2%

ordered simple swap using sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 33.56 19.51 10.5%

1 iter. 36.23 20.15 15.0%

10 iter. 36.93 19.52 14.7%

Protocol Sensitivity Summary for Unidirectional 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 2.5081e-02 2.4493e-05 0.65 0.57 0.60 1.36

16 1.2533e-02 2.4478e-05 1.31 0.58 0.60 1.35

32 6.2976e-03 2.4600e-05 2.60 0.58 0.59 1.35

64 3.7414e-03 2.9230e-05 4.38 0.49 0.49 1.17

128 2.0730e-03 3.2391e-05 7.90 0.46 0.46 1.05

256 1.2468e-03 3.8962e-05 13.14 0.39 0.40 0.92

512 1.1098e-03 6.9362e-05 14.76 0.10 0.05 0.33

1024 7.1340e-04 8.9175e-05 22.97 0.08 0.04 0.28

2048 5.5680e-04 1.3920e-04 29.43 0.07 0.03 0.19

4096 5.0180e-04 2.5090e-04 32.65 0.05 0.04 0.13

8192 4.8700e-04 4.8700e-04 33.64 0.03 0.03 0.08

Five Fastest
Protocols

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

8 1 0 2 3 8

16 0 2 1 3 8

32 0 1 3 2 8

64 1 0 2 3 10

128 1 0 3 2 7

256 0 1 2 3 10

512 1 7 10 3 2

1024 1 0 10 3 2

2048 1 0 7 8 2

4096 0 10 1 2 7

8192 0 10 2 3 8

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 2 4 4

16 1 3 4

32 2 2 4

64 1 3 4

128 1 2 4

256 1 1 4

512 2 6 9

1024 2 7 9

2048 1 7 11

4096 1 8 11

8192 3 9 11

Protocol Sensitivity Summary for Unidirectional 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 2.5110e-02 2.4521e-05 0.65 0.57 0.58 1.35

16 1.2552e-02 2.4515e-05 1.31 0.56 0.57 1.33

32 6.2220e-03 2.4305e-05 2.63 0.57 0.59 1.37

64 3.6404e-03 2.8441e-05 4.50 0.50 0.50 1.17

128 2.0148e-03 3.1481e-05 8.13 0.46 0.46 1.08

256 1.2272e-03 3.8350e-05 13.35 0.37 0.36 0.88

512 1.0676e-03 6.6725e-05 15.35 0.09 0.04 0.31

1024 6.7360e-04 8.4200e-05 24.32 0.08 0.04 0.25

2048 5.3140e-04 1.3285e-04 30.83 0.05 0.01 0.15

4096 4.7820e-04 2.3910e-04 34.26 0.03 0.02 0.09

8192 4.5220e-04 4.5220e-04 36.23 0.02 0.01 0.07

Five Fastest
Protocols

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

8 0 1 2 3 8

16 0 1 2 3 9

32 0 1 3 2 8

64 1 0 3 2 9

128 1 0 2 3 9

256 0 1 2 3 8

512 10 7 1 0 8

1024 10 1 0 7 2

2048 3 1 2 10 8

4096 8 7 1 10 0

8192 10 7 3 1 2

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 2 4 4

16 2 4 4

32 2 4 4

64 2 2 4

128 2 3 4

256 2 3 4

512 3 8 9

1024 2 8 9

2048 3 8 11

4096 2 8 11

8192 2 10 11

Protocol Sensitivity Summary for Unidirectional 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 2.4838e-02 2.4256e-05 0.66 0.57 0.58 1.37

16 1.2404e-02 2.4227e-05 1.32 0.58 0.59 1.38

32 6.1854e-03 2.4162e-05 2.65 0.58 0.58 1.38

64 3.6222e-03 2.8298e-05 4.52 0.51 0.50 1.21

128 2.0074e-03 3.1365e-05 8.16 0.47 0.46 1.08

256 1.1940e-03 3.7314e-05 13.72 0.41 0.41 0.90

512 1.0582e-03 6.6138e-05 15.48 0.09 0.03 0.30

1024 6.6730e-04 8.3413e-05 24.55 0.07 0.04 0.24

2048 5.2318e-04 1.3080e-04 31.32 0.05 0.02 0.14

4096 4.6978e-04 2.3489e-04 34.88 0.02 0.01 0.06

8192 4.4364e-04 4.4364e-04 36.93 0.01 0.00 0.04

Five Fastest
Protocols

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

8 1 0 2 3 8

16 1 0 2 3 8

32 1 0 2 3 8

64 1 0 2 3 9

128 1 0 3 2 9

256 0 1 2 3 9

512 7 0 10 1 2

1024 7 10 1 0 2

2048 10 1 7 0 9

4096 7 0 2 3 1

8192 10 8 0 1 7

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 2 4 4

16 2 4 4

32 2 4 4

64 2 4 4

128 1 2 4

256 1 2 4

512 3 8 9

1024 4 8 11

2048 4 8 11

4096 8 8 11

8192 8 11 11

Protocol Sensitivity Summary for Unidirectional 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 2.5782e-02 2.5178e-05 0.64 0.30 0.02 1.15

16 1.2909e-02 2.5212e-05 1.27 0.30 0.03 1.16

32 6.3958e-03 2.4984e-05 2.56 0.31 0.05 1.17

64 3.7296e-03 2.9138e-05 4.39 0.28 0.06 1.08

128 2.0794e-03 3.2491e-05 7.88 0.28 0.10 1.04

256 1.2746e-03 3.9831e-05 12.85 0.23 0.08 0.80

512 1.1290e-03 7.0563e-05 14.51 0.05 0.02 0.29

1024 7.2240e-04 9.0300e-05 22.68 0.04 0.02 0.21

2048 5.6540e-04 1.4135e-04 28.98 0.03 0.02 0.17

4096 4.9880e-04 2.4940e-04 32.85 0.04 0.03 0.09

8192 4.8100e-04 4.8100e-04 34.06 0.03 0.03 0.06

Five Fastest
Protocols

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

8 2 4 0 1 3

16 0 2 1 4 5

32 0 1 4 2 3

64 0 1 4 2 3

128 0 1 4 2 5

256 0 1 4 2 3

512 2 4 3 0 5

1024 0 2 4 7 1

2048 4 3 8 2 0

4096 4 2 3 5 8

8192 2 3 0 8 4

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 4 6 6

16 2 6 6

32 3 5 6

64 1 5 6

128 1 3 6

256 1 2 6

512 1 10 10

1024 3 9 11

2048 5 10 11

4096 2 8 11

8192 3 9 11

Protocol Sensitivity Summary for Unidirectional 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 2.5220e-02 2.4629e-05 0.65 0.34 0.03 1.24

16 1.2632e-02 2.4673e-05 1.30 0.33 0.03 1.21

32 6.2742e-03 2.4509e-05 2.61 0.34 0.04 1.21

64 3.6810e-03 2.8758e-05 4.45 0.29 0.03 1.10

128 2.0336e-03 3.1775e-05 8.06 0.28 0.05 1.06

256 1.2242e-03 3.8256e-05 13.38 0.24 0.06 0.81

512 1.0968e-03 6.8550e-05 14.94 0.03 0.01 0.22

1024 6.9480e-04 8.6850e-05 23.58 0.03 0.02 0.18

2048 5.4240e-04 1.3560e-04 30.21 0.02 0.01 0.13

4096 4.7720e-04 2.3860e-04 34.33 0.03 0.02 0.09

8192 4.5080e-04 4.5080e-04 36.34 0.02 0.01 0.05

Five Fastest
Protocols

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

8 1 0 2 4 5

16 1 4 0 2 5

32 2 0 1 4 3

64 0 1 4 2 5

128 0 1 4 3 5

256 0 1 4 2 3

512 0 10 7 1 4

1024 0 7 5 1 10

2048 2 3 4 8 10

4096 10 4 5 7 0

8192 5 2 7 9 10

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 3 6 6

16 3 6 6

32 3 6 6

64 2 6 6

128 2 6 6

256 1 3 6

512 3 10 11

1024 4 10 11

2048 5 10 11

4096 1 10 11

8192 3 10 11

Protocol Sensitivity Summary for Unidirectional 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 2.5776e-02 2.5172e-05 0.64 0.40 0.23 1.15

16 1.2901e-02 2.5197e-05 1.27 0.40 0.24 1.15

32 6.4044e-03 2.5017e-05 2.56 0.41 0.23 1.15

64 3.7317e-03 2.9154e-05 4.39 0.36 0.20 1.09

128 2.0642e-03 3.2253e-05 7.94 0.34 0.20 1.00

256 1.2245e-03 3.8267e-05 13.38 0.29 0.18 0.83

512 1.0755e-03 6.7220e-05 15.23 0.07 0.06 0.23

1024 6.7906e-04 8.4882e-05 24.13 0.05 0.06 0.18

2048 5.3124e-04 1.3281e-04 30.84 0.03 0.03 0.08

4096 4.7064e-04 2.3532e-04 34.81 0.01 0.01 0.04

8192 4.4000e-04 4.4000e-04 37.24 0.01 0.01 0.03

Five Fastest
Protocols

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

8 0 1 4 2 5

16 0 1 4 2 5

32 0 1 4 2 5

64 0 1 3 4 5

128 0 1 2 4 3

256 0 1 2 4 5

512 10 0 1 7 2

1024 1 7 0 10 2

2048 7 1 10 0 8

4096 4 8 2 0 10

8192 4 8 3 2 5

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 2 2 6

16 2 2 6

32 2 2 6

64 2 2 6

128 2 2 6

256 2 2 6

512 2 4 11

1024 4 4 11

2048 3 10 11

4096 9 11 11

8192 6 11 11

DISCUSSION

The AlphaServer SC 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 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.