PSTSWM AlphaSC-500 Point-to-Point Communication Performance

Performance Studies using

PSTSWM

Compaq AlphaServer SC SENDRECV Performance

(ordered sendrecv of 128KB message using MPI between two nodes)

(performance measured per processor when all processors send right, read left in a logical ring)

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:	16384 REAL*8 floating point values each direction
Message size:	Largest - 16384 REAL*8 floating point values
	Smallest - 16 REAL*8 floating point values
Processors:	0 -> 1 -> 2 -> 3 -> 4 -> 5 -> 6 -> 7 -> 0
Latency Definition:	(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)

1 iter. 115.89 11.78 65.0%

10 iter. 134.54 16.83 54.4%

ordered swap using nonblocking send

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 113.91 13.99 59.2%

10 iter. 139.14 16.59 54.1%

1 iter. w/overlap 113.94 14.69 58.7%

10 iter. w/overlap 141.76 17.71 54.7%

ordered swap using nonblocking receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 112.82 19.91 47.6%

10 iter. 140.45 21.24 45.0%

1 iter. w/overlap 119.18 13.99 62.2%

10 iter. w/overlap 141.68 16.31 59.4%

ordered swap using nonblocking send and receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 115.85 26.00 35.5%

10 iter. 139.53 23.06 42.0%

1 iter. w/overlap 121.93 14.96 60.4%

10 iter. w/overlap 144.16 16.30 58.9%

ordered swap using ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 112.83 36.84 40.0%

10 iter. 133.00 36.65 40.7%

1 iter. w/overlap 119.11 39.49 38.9%

10 iter. w/overlap 140.59 55.77 30.7%

ordered swap using nonblocking ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 112.03 36.24 38.9%

10 iter. 137.54 35.92 39.5%

1 iter. w/overlap 118.11 38.39 39.2%

10 iter. w/overlap 138.05 55.13 30.7%

synchronous

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 112.62 38.92 35.8%

10 iter. 138.84 38.58 37.3%

ordered swap using nonblocking sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 118.47 23.86 45.5%

10 iter. 149.96 23.55 46.9%

1 iter. w/overlap 121.17 24.33 40.2%

10 iter. w/overlap 145.43 22.92 46.2%

ordered swap using nonblocking receive with sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 117.21 27.59 37.8%

10 iter. 145.18 30.46 33.0%

1 iter. w/overlap 119.73 19.13 53.4%

10 iter. w/overlap 141.75 22.27 50.8%

ordered swap using nonblocking sync. send and receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 122.10 30.31 33.2%

10 iter. 149.23 29.73 34.0%

1 iter. w/overlap 119.86 17.94 55.3%

10 iter. w/overlap 142.85 21.63 50.0%

ordered simple swap using sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

1 iter. 115.11 22.91 49.6%

10 iter. 138.24 21.00 54.2%

Protocol Sensitivity Summary for Unidirectional Swap of 131072 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

128 3.6892e-02 3.6027e-05 7.11 0.71 0.65 1.45

256 2.3988e-02 4.6851e-05 10.93 0.51 0.52 1.11

512 2.3187e-02 9.0575e-05 11.31 0.20 0.14 0.52

1024 1.2752e-02 9.9625e-05 20.56 0.18 0.10 0.44

2048 7.4024e-03 1.1566e-04 35.41 0.12 0.05 0.35

4096 4.6302e-03 1.4469e-04 56.62 0.10 0.05 0.28

8192 3.2290e-03 2.0181e-04 81.18 0.08 0.05 0.22

16384 2.5710e-03 3.2137e-04 101.96 0.06 0.05 0.17

32768 2.2778e-03 5.6945e-04 115.09 0.06 0.03 0.17

65536 2.1470e-03 1.0735e-03 122.10 0.07 0.07 0.16

131072 2.2628e-03 2.2628e-03 115.85 0.02 0.02 0.03

Five Fastest
Protocols

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

128 0 1 2 3 10

256 1 0 3 2 10

512 1 0 2 3 8

1024 0 1 2 3 9

2048 0 1 7 9 3

4096 3 2 0 1 9

8192 1 3 2 7 9

16384 8 9 7 3 1

32768 8 7 3 1 2

65536 9 7 8 0 10

131072 3 9 10 0 7

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

128 1 2 2

256 1 2 4

512 1 4 8

1024 1 3 8

2048 2 5 8

4096 3 4 9

8192 1 4 11

16384 3 5 11

32768 1 7 11

65536 1 3 11

131072 4 11 11

Protocol Sensitivity Summary for Unidirectional Swap of 131072 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

128 4.1314e-02 4.0346e-05 6.35 0.55 0.48 1.19

256 2.4331e-02 4.7521e-05 10.77 0.51 0.53 1.10

512 2.2617e-02 8.8350e-05 11.59 0.21 0.15 0.54

1024 1.2406e-02 9.6920e-05 21.13 0.17 0.10 0.45

2048 7.2260e-03 1.1291e-04 36.28 0.13 0.09 0.35

4096 4.5680e-03 1.4275e-04 57.39 0.09 0.06 0.25

8192 3.1714e-03 1.9821e-04 82.66 0.07 0.04 0.21

16384 2.4900e-03 3.1125e-04 105.28 0.06 0.04 0.15

32768 2.1793e-03 5.4482e-04 120.29 0.03 0.01 0.08

65536 1.8666e-03 9.3332e-04 140.44 0.05 0.06 0.11

131072 1.7481e-03 1.7481e-03 149.96 0.07 0.08 0.13

Five Fastest
Protocols

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

128 1 0 2 3 10

256 1 0 2 3 10

512 1 2 0 3 9

1024 1 0 3 2 9

2048 0 1 3 2 8

4096 2 1 0 3 9

8192 1 0 3 2 7

16384 7 8 9 10 1

32768 9 7 1 2 8

65536 8 7 9 10 1

131072 7 9 8 2 3

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

128 1 1 3

256 1 1 4

512 1 4 8

1024 2 4 8

2048 1 4 8

4096 2 4 10

8192 1 8 11

16384 2 8 11

32768 4 7 11

65536 1 4 11

131072 2 3 11

Protocol Sensitivity Summary for Unidirectional Swap of 131072 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

128 3.7843e-02 3.6956e-05 6.93 0.52 0.27 1.38

256 2.2799e-02 4.4529e-05 11.50 0.46 0.24 1.19

512 2.3624e-02 9.2280e-05 11.10 0.19 0.08 0.55

1024 1.3105e-02 1.0238e-04 20.00 0.15 0.05 0.45

2048 7.4618e-03 1.1659e-04 35.13 0.14 0.08 0.41

4096 4.6018e-03 1.4381e-04 56.97 0.12 0.10 0.32

8192 3.3070e-03 2.0669e-04 79.27 0.06 0.02 0.21

16384 2.5382e-03 3.1728e-04 103.28 0.08 0.07 0.17

32768 2.2532e-03 5.6330e-04 116.34 0.07 0.05 0.33

65536 2.1500e-03 1.0750e-03 121.93 0.06 0.02 0.33

131072 2.2602e-03 2.2602e-03 115.98 0.02 0.02 0.05

Five Fastest
Protocols

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

128 1 2 3 0 9

256 1 2 3 0 9

512 0 1 7 3 2

1024 1 2 0 3 7

2048 1 2 3 0 7

4096 1 2 0 9 7

8192 2 1 8 7 0

16384 3 7 2 0 1

32768 3 2 9 1 7

65536 3 7 9 8 2

131072 3 9 7 8 4

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

128 1 3 5

256 1 2 6

512 2 5 8

1024 4 6 8

2048 1 5 8

4096 1 4 8

8192 3 8 11

16384 2 2 11

32768 3 6 10

65536 2 8 10

131072 4 11 11

Protocol Sensitivity Summary for Unidirectional Swap of 131072 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

128 4.1073e-02 4.0110e-05 6.38 0.60 0.31 1.96

256 2.4370e-02 4.7599e-05 10.76 0.54 0.27 1.67

512 2.3771e-02 9.2854e-05 11.03 0.25 0.09 0.85

1024 1.2689e-02 9.9133e-05 20.66 0.23 0.08 0.78

2048 7.3632e-03 1.1505e-04 35.60 0.18 0.05 0.66

4096 4.6308e-03 1.4471e-04 56.61 0.13 0.03 0.47

8192 3.2253e-03 2.0158e-04 81.28 0.09 0.03 0.32

16384 2.5526e-03 3.1907e-04 102.70 0.05 0.01 0.17

32768 2.1577e-03 5.3942e-04 121.49 0.03 0.01 0.09

65536 1.9024e-03 9.5121e-04 137.80 0.03 0.02 0.10

131072 1.8025e-03 1.8025e-03 145.43 0.04 0.03 0.16

Five Fastest
Protocols

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

128 2 3 0 1 9

256 2 3 1 0 9

512 0 1 3 2 7

1024 1 0 2 3 7

2048 0 1 2 3 7

4096 0 2 1 8 7

8192 0 1 2 7 3

16384 2 1 7 3 9

32768 9 3 2 7 1

65536 7 3 8 2 9

131072 7 10 3 9 1

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

128 1 2 4

256 1 2 5

512 1 3 8

1024 2 3 8

2048 1 6 8

4096 2 7 9

8192 2 7 9

16384 4 8 11

32768 5 7 11

65536 5 8 11

131072 3 8 11

DISCUSSION

The AlphaServer SC processing node consists of four processors. In these tests, we examine MPI communication performance between processors in different nodes when only nearest neighbor communication is required. The performance is reported for a single processor pair (logical processor 0), representing the latency and bandwdith that would typically appear in a nearest neighbor algorithm, such as a 1D halo exchange.