PSTSWM AlphaSC-500 Point-to-Point Communication Performance

Performance Studies using

PSTSWM

Compaq AlphaServer SC SWAP Performance

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

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

Date/Person:	January 26, 2000 / P. Worley
Platform:	Compaq AlphaServer SC at Oak Ridge National Laboratory (colt.ccs.ornl.gov):
	16 ES40 4-way SMP nodes (500 MHz Alpha 21264 with 4MB L2 cache)
Environment:	Digital UNIX V5.0; RMS 2.36
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
	1 and 5
	2 and 6
	3 and 7
Latency Definition:	(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)

1 iter. 38.97 47.93 22.8%

10 iter. 35.75 47.07 18.9%

unordered swap using nonblocking send

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 35.98 44.49 31.8%

10 iter. 37.47 44.67 28.6%

1 iter. w/overlap 35.19 45.03 28.2%

10 iter. w/overlap 36.70 40.94 14.9%

unordered swap using nonblocking receive

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 37.56 45.96 59.3%

10 iter. 37.36 45.88 29.1%

1 iter. w/overlap 38.41 21.38 21.4%

10 iter. w/overlap 37.97 23.57 21.5%

unordered swap using nonblocking send and receive

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 40.08 46.37 26.4%

10 iter. 37.56 46.85 30.9%

1 iter. w/overlap 35.60 20.65 17.9%

10 iter. w/overlap 37.67 23.80 17.0%

unordered swap using ready send

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 31.84 68.54 29.3%

10 iter. 35.52 67.36 22.7%

1 iter. w/overlap 39.50 20.55 23.6%

10 iter. w/overlap 38.51 23.41 21.9%

unordered swap using nonblocking ready send

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 35.02 68.31 29.2%

10 iter. 35.48 69.97 23.5%

1 iter. w/overlap 40.61 20.24 23.0%

10 iter. w/overlap 37.48 23.44 16.8%

native sendrecv

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 35.74 45.80 40.8%

10 iter. 37.05 45.71 30.3%

unordered swap using nonblocking sync. send

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 37.70 34.94 16.1%

10 iter. 36.85 35.02 16.1%

1 iter. w/overlap 35.68 46.09 20.1%

10 iter. w/overlap 36.41 39.81 23.2%

unordered swap using nonblocking receive with sync. send

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 34.90 36.17 21.6%

10 iter. 38.05 36.32 16.9%

1 iter. w/overlap 35.43 36.10 17.0%

10 iter. w/overlap 38.36 35.83 12.9%

unordered swap using nonblocking sync. send and receive

Data Statistics
bidirectional bandwidth estimated latency model error

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

1 iter. 40.37 36.12 12.5%

10 iter. 37.50 36.47 17.0%

1 iter. w/overlap 34.92 33.45 26.4%

10 iter. w/overlap 37.72 31.59 9.8%

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 3.5828e-02 3.4989e-05 0.46 0.35 0.31 0.93

16 1.7939e-02 3.5037e-05 0.91 0.33 0.30 0.90

32 8.9748e-03 3.5058e-05 1.83 0.33 0.30 0.87

64 5.1358e-03 4.0123e-05 3.19 0.30 0.27 0.81

128 2.8410e-03 4.4391e-05 5.77 0.22 0.16 0.67

256 1.6206e-03 5.0644e-05 10.11 0.19 0.16 0.59

512 7.3540e-04 4.5963e-05 22.28 0.45 0.35 0.97

1024 5.8580e-04 7.3225e-05 27.97 0.15 0.11 0.43

2048 5.0520e-04 1.2630e-04 32.43 0.10 0.08 0.25

4096 4.0880e-04 2.0440e-04 40.08 0.12 0.14 0.26

8192 4.0580e-04 4.0580e-04 40.37 0.18 0.18 0.31

Five Fastest
Protocols

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

8 7 9 8 1 6

16 7 9 8 0 1

32 7 9 8 0 1

64 7 8 9 6 3

128 7 8 9 2 3

256 3 8 9 7 0

512 2 8 7 1 6

1024 6 1 3 2 8

2048 8 3 1 2 6

4096 3 0 7 2 6

8192 9 1 6 3 8

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 3 3

16 1 3 3

32 1 3 4

64 1 3 4

128 1 4 7

256 1 3 8

512 1 1 1

1024 1 1 8

2048 4 5 10

4096 1 2 9

8192 1 1 8

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 3.5954e-02 3.5111e-05 0.46 0.35 0.31 0.96

16 1.8025e-02 3.5205e-05 0.91 0.34 0.31 0.93

32 8.9345e-03 3.4900e-05 1.83 0.36 0.32 0.97

64 5.1655e-03 4.0355e-05 3.17 0.32 0.29 0.83

128 2.8401e-03 4.4376e-05 5.77 0.26 0.22 0.71

256 1.6814e-03 5.2543e-05 9.74 0.19 0.16 0.58

512 9.5066e-04 5.9416e-05 17.23 0.13 0.02 0.48

1024 6.1774e-04 7.7217e-05 26.52 0.09 0.01 0.34

2048 4.9538e-04 1.2385e-04 33.07 0.06 0.02 0.21

4096 4.3054e-04 2.1527e-04 38.05 0.08 0.07 0.17

8192 4.3620e-04 4.3620e-04 37.56 0.02 0.01 0.06

Five Fastest
Protocols

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

8 7 8 9 1 6

16 7 8 9 1 2

32 7 8 9 1 0

64 7 9 8 3 6

128 7 8 9 0 3

256 7 8 9 0 3

512 1 9 6 2 7

1024 1 8 3 6 9

2048 3 9 1 8 2

4096 8 9 1 3 2

8192 3 9 1 8 2

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 3 3

16 1 3 3

32 1 3 3

64 1 3 3

128 1 3 7

256 2 3 8

512 3 7 8

1024 6 7 8

2048 2 7 10

4096 1 1 10

8192 5 7 10

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 2.1317e-02 2.0818e-05 0.77 0.62 0.73 1.24

16 1.0708e-02 2.0915e-05 1.53 0.58 0.72 1.25

32 5.2790e-03 2.0621e-05 3.10 0.59 0.71 1.33

64 3.2194e-03 2.5152e-05 5.09 0.52 0.62 1.20

128 1.8470e-03 2.8859e-05 8.87 0.40 0.46 0.95

256 1.1406e-03 3.5644e-05 14.36 0.39 0.43 1.07

512 9.4140e-04 5.8838e-05 17.40 0.07 0.04 0.24

1024 6.2920e-04 7.8650e-05 26.04 0.04 0.03 0.15

2048 4.7700e-04 1.1925e-04 34.35 0.09 0.09 0.25

4096 4.0340e-04 2.0170e-04 40.61 0.14 0.14 0.28

8192 3.7840e-04 3.7840e-04 43.30 0.21 0.24 0.32

Five Fastest
Protocols

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

8 4 3 5 2 9

16 2 3 4 5 9

32 2 3 4 5 9

64 2 5 4 3 9

128 3 4 5 2 0

256 2 4 5 3 9

512 9 5 3 4 2

1024 9 4 5 3 1

2048 9 3 5 6 2

4096 5 6 4 2 7

8192 6 4 2 8 0

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 2 4 4

16 3 4 4

32 1 3 4

64 1 4 4

128 2 4 4

256 1 4 4

512 2 7 10

1024 2 8 10

2048 1 3 10

4096 1 1 9

8192 1 1 6

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 2.3905e-02 2.3345e-05 0.69 0.43 0.53 0.97

16 1.1918e-02 2.3278e-05 1.37 0.43 0.54 0.97

32 5.9518e-03 2.3249e-05 2.75 0.43 0.53 0.96

64 3.4328e-03 2.6818e-05 4.77 0.47 0.54 0.94

128 1.9606e-03 3.0634e-05 8.36 0.40 0.48 0.89

256 1.2106e-03 3.7831e-05 13.53 0.33 0.41 0.85

512 9.7638e-04 6.1024e-05 16.78 0.06 0.05 0.20

1024 6.3648e-04 7.9560e-05 25.74 0.05 0.04 0.12

2048 5.0536e-04 1.2634e-04 32.42 0.04 0.02 0.11

4096 4.5278e-04 2.2639e-04 36.19 0.02 0.01 0.07

8192 4.2548e-04 4.2548e-04 38.51 0.03 0.03 0.08

Five Fastest
Protocols

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

8 4 2 5 3 9

16 4 2 3 5 9

32 2 4 3 5 9

64 4 2 3 5 9

128 2 4 5 3 9

256 2 5 3 4 9

512 5 3 9 6 8

1024 6 3 5 8 2

2048 8 5 9 6 3

4096 5 9 4 3 8

8192 4 8 2 9 3

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

8 1 4 4

16 1 4 4

32 1 4 4

64 1 4 4

128 1 4 4

256 2 4 4

512 2 6 10

1024 2 7 10

2048 1 8 10

4096 4 8 10

8192 2 8 10

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 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.