PSTSWM AlphaSC-1000 Point-to-Point Communication Performance

Performance Studies using

PSTSWM

Compaq AlphaServer SC SWAP Performance

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

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

Date/Person:	January 3, 2002 / P. Worley
Platform:	Compaq AlphaServer SC at Pittsburgh Supercomputer Center (lemieux.psc.edu):
	750 ES45 4-way SMP nodes (1.0 GHz Alpha 21264C with 8MB L2 cache)
Environment:	SC System TS2.5; Digital UNIX V5.1; RMS X.XX
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:	0 and 1
	2 and 3
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)

cache inv. 424.62 26.75 35.6%

1 iter. 555.60 26.66 34.4%

10 iter. 587.79 26.73 39.1%

ordered swap using nonblocking send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 449.48 26.65 27.8%

1 iter. 522.32 26.86 36.8%

10 iter. 586.10 26.76 35.6%

cache inv. w/overlap 445.97 27.35 27.3%

1 iter. w/overlap 525.97 27.16 33.6%

10 iter. w/overlap 583.73 28.69 35.5%

ordered swap using nonblocking receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 437.66 25.80 22.3%

1 iter. 548.42 26.23 31.2%

10 iter. 589.51 26.67 31.4%

cache inv. w/overlap 443.84 26.58 22.8%

1 iter. w/overlap 552.89 26.80 33.3%

10 iter. w/overlap 585.19 29.29 36.6%

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. 427.74 25.37 25.0%

1 iter. 552.07 26.70 26.7%

10 iter. 587.69 26.56 29.7%

cache inv. w/overlap 442.53 27.61 24.3%

1 iter. w/overlap 554.67 26.33 17.3%

10 iter. w/overlap 584.55 29.80 29.8%

ordered swap using ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 449.13 37.87 26.7%

1 iter. 509.67 40.39 39.5%

10 iter. 582.69 41.28 39.5%

cache inv. w/overlap 442.51 27.05 29.2%

1 iter. w/overlap 517.11 26.41 33.6%

10 iter. w/overlap 586.15 29.30 36.3%

ordered swap using nonblocking ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 439.16 36.79 21.4%

1 iter. 551.50 38.05 27.8%

10 iter. 583.05 39.33 33.2%

cache inv. w/overlap 439.30 26.61 19.5%

1 iter. w/overlap 553.73 26.28 22.7%

10 iter. w/overlap 586.51 29.83 29.3%

synchronous

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 426.67 39.45 40.7%

1 iter. 552.45 42.29 47.7%

10 iter. 586.27 42.92 45.3%

ordered swap using nonblocking sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 450.68 31.92 40.9%

1 iter. 517.97 32.00 45.9%

10 iter. 584.13 32.04 41.2%

cache inv. w/overlap 442.72 32.26 38.9%

1 iter. w/overlap 514.35 31.58 43.7%

10 iter. w/overlap 582.71 33.37 37.2%

ordered swap using nonblocking receive with sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 449.32 31.65 31.0%

1 iter. 555.60 31.84 38.7%

10 iter. 585.86 31.87 39.1%

cache inv. w/overlap 446.87 32.37 37.4%

1 iter. w/overlap 555.68 31.82 38.7%

10 iter. w/overlap 587.73 33.97 37.0%

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. 426.36 31.80 38.9%

1 iter. 556.01 32.04 36.6%

10 iter. 585.79 31.95 38.7%

cache inv. w/overlap 427.03 32.41 37.0%

1 iter. w/overlap 550.09 32.07 32.5%

10 iter. w/overlap 584.19 33.47 30.3%

ordered simple swap using sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 449.38 31.80 42.9%

1 iter. 526.41 31.93 45.9%

10 iter. 586.72 31.81 41.9%

Protocol Sensitivity Summary for Unidirectional 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 5.5762e-02 5.4455e-05 75.22 0.17 0.17 0.42

4096 2.8368e-02 5.5407e-05 147.85 0.17 0.16 0.37

8192 1.7389e-02 6.7927e-05 241.20 0.12 0.08 0.35

16384 1.4403e-02 1.1252e-04 291.21 0.06 0.05 0.15

32768 1.3340e-02 2.0843e-04 314.42 0.03 0.03 0.07

65536 1.1908e-02 3.7211e-04 352.24 0.02 0.02 0.03

131072 1.0806e-02 6.7537e-04 388.15 0.01 0.01 0.03

262144 1.0074e-02 1.2592e-03 416.36 0.01 0.01 0.01

524288 9.7844e-03 2.4461e-03 428.67 0.02 0.01 0.06

1048576 9.5258e-03 4.7629e-03 440.31 0.02 0.01 0.04

2097152 9.3992e-03 9.3992e-03 446.24 0.04 0.03 0.09

Five Fastest
Protocols

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

2048 0 2 1 3 10

4096 0 1 2 3 10

8192 0 10 1 7 3

16384 0 10 6 3 9

32768 0 2 10 7 3

65536 0 6 2 3 10

131072 0 10 7 2 6

262144 1 10 2 8 7

524288 7 4 1 10 8

1048576 10 1 2 7 4

2097152 2 6 7 8 4

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 4 4 8

4096 1 3 8

8192 1 5 9

16384 2 6 11

32768 1 9 11

65536 3 11 11

131072 5 11 11

262144 8 11 11

524288 7 7 11

1048576 4 11 11

2097152 2 8 11

Protocol Sensitivity Summary for Unidirectional 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 5.5449e-02 5.4149e-05 75.64 0.19 0.18 0.48

4096 2.8153e-02 5.4987e-05 148.98 0.17 0.16 0.39

8192 1.5773e-02 6.1614e-05 265.91 0.12 0.06 0.34

16384 1.2370e-02 9.6642e-05 339.07 0.08 0.05 0.15

32768 1.1312e-02 1.7675e-04 370.79 0.05 0.05 0.10

65536 9.7834e-03 3.0573e-04 428.72 0.03 0.03 0.05

131072 8.7162e-03 5.4476e-04 481.21 0.01 0.01 0.02

262144 8.1238e-03 1.0155e-03 516.30 0.02 0.01 0.05

524288 7.8334e-03 1.9584e-03 535.44 0.01 0.01 0.05

1048576 7.6374e-03 3.8187e-03 549.18 0.04 0.01 0.09

2097152 7.5768e-03 7.5768e-03 553.57 0.02 0.01 0.11

Five Fastest
Protocols

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

2048 0 2 1 3 10

4096 0 1 2 3 10

8192 0 1 2 10 7

16384 10 1 2 7 8

32768 2 1 10 8 7

65536 0 2 8 3 10

131072 0 10 2 7 1

262144 0 3 2 9 6

524288 2 8 3 0 9

1048576 0 3 9 5 8

2097152 4 3 10 9 0

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 3 4 8

4096 1 4 8

8192 2 3 8

16384 1 5 11

32768 2 6 11

65536 1 11 11

131072 3 11 11

262144 5 10 11

524288 6 10 11

1048576 6 7 11

2097152 6 10 11

Protocol Sensitivity Summary for Unidirectional 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 5.5326e-02 5.4029e-05 75.81 0.19 0.18 0.50

4096 2.7956e-02 5.4602e-05 150.03 0.17 0.17 0.39

8192 1.4968e-02 5.8468e-05 280.22 0.15 0.12 0.37

16384 1.1206e-02 8.7549e-05 374.28 0.07 0.05 0.19

32768 1.0099e-02 1.5780e-04 415.31 0.04 0.04 0.10

65536 8.9197e-03 2.7874e-04 470.23 0.02 0.02 0.04

131072 7.8561e-03 4.9100e-04 533.89 0.01 0.01 0.03

262144 7.3604e-03 9.2005e-04 569.85 0.01 0.01 0.02

524288 7.1487e-03 1.7872e-03 586.72 0.01 0.01 0.02

1048576 7.1149e-03 3.5575e-03 589.51 0.01 0.01 0.02

2097152 7.3470e-03 7.3470e-03 570.89 0.01 0.02 0.03

Five Fastest
Protocols

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

2048 0 2 1 3 10

4096 0 1 2 3 10

8192 0 1 2 3 10

16384 0 10 2 1 8

32768 0 3 10 9 6

65536 2 0 10 8 3

131072 0 10 1 3 6

262144 8 2 0 3 9

524288 10 1 7 4 0

1048576 2 0 3 6 8

2097152 4 3 5 2 1

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 4 4 8

4096 1 4 8

8192 1 2 8

16384 1 6 11

32768 1 8 11

65536 3 11 11

131072 6 11 11

262144 5 11 11

524288 4 11 11

1048576 6 11 11

2097152 4 11 11

Protocol Sensitivity Summary for Unidirectional 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 5.7038e-02 5.5701e-05 73.54 0.11 0.03 0.44

4096 2.9070e-02 5.6778e-05 144.28 0.11 0.06 0.37

8192 1.7122e-02 6.6884e-05 244.96 0.10 0.10 0.23

16384 1.4089e-02 1.1007e-04 297.69 0.08 0.09 0.17

32768 1.3251e-02 2.0705e-04 316.52 0.04 0.03 0.08

65536 1.1759e-02 3.6748e-04 356.68 0.03 0.03 0.06

131072 1.0552e-02 6.5951e-04 397.48 0.02 0.02 0.04

262144 1.0003e-02 1.2504e-03 419.29 0.01 0.01 0.02

524288 9.7452e-03 2.4363e-03 430.40 0.03 0.01 0.08

1048576 9.5210e-03 4.7605e-03 440.53 0.03 0.01 0.08

2097152 9.4700e-03 9.4700e-03 442.90 0.03 0.01 0.09

Five Fastest
Protocols

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

2048 0 2 1 4 5

4096 0 1 2 4 3

8192 0 10 4 1 7

16384 0 10 4 6 1

32768 10 0 4 6 8

65536 0 10 4 1 8

131072 10 4 2 1 5

262144 4 10 5 8 2

524288 10 1 8 2 5

1048576 4 7 2 5 8

2097152 8 6 4 3 5

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 3 6 10

4096 1 5 10

8192 1 2 11

16384 1 3 11

32768 2 8 11

65536 3 10 11

131072 2 11 11

262144 5 11 11

524288 6 7 11

1048576 6 7 11

2097152 3 8 11

Protocol Sensitivity Summary for Unidirectional 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 5.5161e-02 5.3868e-05 76.04 0.12 0.02 0.50

4096 2.8005e-02 5.4696e-05 149.77 0.11 0.05 0.40

8192 1.5694e-02 6.1305e-05 267.26 0.09 0.08 0.26

16384 1.2348e-02 9.6470e-05 339.67 0.07 0.08 0.17

32768 1.1082e-02 1.7315e-04 378.49 0.07 0.05 0.15

65536 9.7916e-03 3.0599e-04 428.36 0.02 0.02 0.03

131072 8.6300e-03 5.3938e-04 486.01 0.03 0.03 0.04

262144 8.1746e-03 1.0218e-03 513.09 0.02 0.01 0.05

524288 7.8582e-03 1.9645e-03 533.75 0.02 0.01 0.09

1048576 7.6866e-03 3.8433e-03 545.66 0.02 0.01 0.07

2097152 7.5800e-03 7.5800e-03 553.34 0.02 0.00 0.11

Five Fastest
Protocols

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

2048 0 4 2 1 3

4096 0 1 4 2 3

8192 0 2 4 1 10

16384 10 2 4 8 0

32768 10 4 8 2 1

65536 0 2 10 8 4

131072 0 8 4 2 9

262144 0 8 2 3 9

524288 0 8 6 2 9

1048576 0 8 9 6 2

2097152 4 0 1 5 3

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 2 6 10

4096 1 4 10

8192 1 4 10

16384 1 5 11

32768 1 6 11

65536 4 11 11

131072 1 11 11

262144 6 11 11

524288 7 8 11

1048576 7 8 11

2097152 8 9 11

Protocol Sensitivity Summary for Unidirectional 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 5.8919e-02 5.7538e-05 71.19 0.12 0.06 0.46

4096 2.9532e-02 5.7681e-05 142.02 0.12 0.09 0.37

8192 1.5371e-02 6.0045e-05 272.86 0.12 0.12 0.33

16384 1.1131e-02 8.6963e-05 376.81 0.07 0.09 0.14

32768 9.9701e-03 1.5578e-04 420.69 0.05 0.05 0.09

65536 8.8709e-03 2.7721e-04 472.82 0.02 0.02 0.05

131072 7.8311e-03 4.8945e-04 535.59 0.02 0.02 0.03

262144 7.3509e-03 9.1886e-04 570.59 0.01 0.01 0.02

524288 7.1464e-03 1.7866e-03 586.91 0.01 0.02 0.03

1048576 7.1153e-03 3.5576e-03 589.48 0.01 0.01 0.02

2097152 7.3927e-03 7.3927e-03 567.36 0.01 0.01 0.02

Five Fastest
Protocols

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

2048 0 1 2 4 3

4096 0 1 2 4 3

8192 0 1 2 4 10

16384 0 10 2 4 8

32768 0 10 4 2 6

65536 2 8 0 10 4

131072 0 10 4 1 7

262144 2 8 0 5 4

524288 10 4 1 7 8

1048576 0 8 5 6 2

2097152 5 4 2 3 10

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 1 4 10

4096 1 4 10

8192 1 1 10

16384 2 5 11

32768 1 5 11

65536 3 10 11

131072 3 11 11

262144 4 11 11

524288 4 11 11

1048576 7 11 11

2097152 4 11 11

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.