PSTSWM AlphaSC-667 Point-to-Point Communication Performance

Performance Studies using

PSTSWM

Compaq AlphaServer SC SWAP Performance

(ordered swap of 2MB message using MPI between two nodes)

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

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:	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 4 nodes
	1 and 5
	2 and 6
	3 and 7
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. 39.78 13.80 4.7%

1 iter. 39.50 14.04 4.6%

10 iter. 40.64 14.06 3.6%

ordered swap using nonblocking send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 39.19 13.18 2.1%

1 iter. 39.22 12.70 1.3%

10 iter. 40.11 13.10 2.4%

cache inv. w/overlap 39.07 13.80 3.5%

1 iter. w/overlap 39.19 14.08 2.6%

10 iter. w/overlap 39.50 14.75 0.9%

ordered swap using nonblocking receive

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 39.46 14.79 2.5%

1 iter. 39.83 15.17 4.8%

10 iter. 39.71 15.15 1.1%

cache inv. w/overlap 39.15 16.80 3.3%

1 iter. w/overlap 39.26 17.03 2.7%

10 iter. w/overlap 39.55 17.36 3.8%

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. 39.54 15.68 2.0%

1 iter. 39.20 15.78 1.6%

10 iter. 40.54 15.62 3.0%

cache inv. w/overlap 39.32 16.77 3.8%

1 iter. w/overlap 39.12 16.77 3.3%

10 iter. w/overlap 40.22 19.50 3.6%

ordered swap using ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 39.70 22.65 3.1%

1 iter. 39.09 21.10 2.9%

10 iter. 39.53 19.83 4.4%

cache inv. w/overlap 39.32 16.74 2.7%

1 iter. w/overlap 39.31 17.30 3.2%

10 iter. w/overlap 39.44 17.35 4.0%

ordered swap using nonblocking ready send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 39.29 22.43 3.1%

1 iter. 39.45 22.05 2.7%

10 iter. 39.92 21.26 3.8%

cache inv. w/overlap 39.15 16.98 4.2%

1 iter. w/overlap 39.54 17.29 3.8%

10 iter. w/overlap 39.69 18.91 4.1%

synchronous

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 39.55 21.33 7.5%

1 iter. 39.24 22.55 3.1%

10 iter. 39.52 22.66 1.9%

ordered swap using nonblocking sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 39.11 13.02 1.8%

1 iter. 39.23 13.61 2.1%

10 iter. 40.28 13.08 2.7%

cache inv. w/overlap 39.14 13.89 3.6%

1 iter. w/overlap 39.08 13.98 3.7%

10 iter. w/overlap 40.57 14.66 3.4%

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. 39.41 15.66 3.4%

1 iter. 39.50 15.24 3.4%

10 iter. 40.52 14.86 2.8%

cache inv. w/overlap 39.26 17.09 3.0%

1 iter. w/overlap 39.20 17.39 3.3%

10 iter. w/overlap 40.17 17.28 2.3%

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. 39.15 15.93 3.3%

1 iter. 39.12 15.11 1.8%

10 iter. 40.43 15.20 3.2%

cache inv. w/overlap 39.15 16.64 3.8%

1 iter. w/overlap 39.11 16.77 2.5%

10 iter. w/overlap 40.18 19.43 3.7%

ordered simple swap using sync. send

Data Statistics
unidirectional bandwidth estimated latency model error

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

cache inv. 39.55 13.68 3.1%

1 iter. 39.58 13.58 4.4%

10 iter. 39.99 13.84 2.1%

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 1.3386e-01 1.3072e-04 31.33 0.05 0.03 0.15

4096 1.2053e-01 2.3540e-04 34.80 0.02 0.01 0.09

8192 1.1379e-01 4.4448e-04 36.86 0.01 0.00 0.05

16384 1.1020e-01 8.6091e-04 38.06 0.01 0.01 0.01

32768 1.0653e-01 1.6646e-03 39.37 0.01 0.01 0.02

65536 1.0544e-01 3.2949e-03 39.78 0.01 0.02 0.02

131072 1.0566e-01 6.6039e-03 39.70 0.01 0.01 0.02

262144 1.0646e-01 1.3307e-02 39.40 0.01 0.01 0.01

524288 1.0629e-01 2.6573e-02 39.46 0.01 0.01 0.02

1048576 1.0827e-01 5.4137e-02 38.74 0.01 0.01 0.02

2097152 1.0766e-01 1.0766e-01 38.96 0.01 0.01 0.07

Five Fastest
Protocols

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

2048 7 1 0 10 2

4096 7 1 0 10 9

8192 7 10 0 1 3

16384 3 8 9 2 7

32768 10 0 3 2 1

65536 0 6 8 4 9

131072 4 10 6 3 8

262144 8 5 3 9 1

524288 2 10 8 5 0

1048576 3 4 2 7 1

2097152 3 2 7 10 5

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 4 8 11

4096 7 8 11

8192 8 10 11

16384 9 11 11

32768 3 11 11

65536 3 11 11

131072 5 11 11

262144 8 11 11

524288 4 11 11

1048576 5 11 11

2097152 5 10 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 1.3404e-01 1.3090e-04 31.29 0.05 0.02 0.15

4096 1.2057e-01 2.3548e-04 34.79 0.03 0.01 0.09

8192 1.1376e-01 4.4437e-04 36.87 0.01 0.00 0.05

16384 1.1023e-01 8.6115e-04 38.05 0.00 0.01 0.01

32768 1.0619e-01 1.6592e-03 39.50 0.02 0.02 0.03

65536 1.0692e-01 3.3412e-03 39.23 0.00 0.00 0.01

131072 1.0631e-01 6.6443e-03 39.45 0.01 0.01 0.01

262144 1.0530e-01 1.3163e-02 39.83 0.01 0.02 0.02

524288 1.0718e-01 2.6794e-02 39.14 0.00 0.00 0.01

1048576 1.0795e-01 5.3974e-02 38.86 0.02 0.01 0.03

2097152 1.0773e-01 1.0773e-01 38.93 0.01 0.00 0.03

Five Fastest
Protocols

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

2048 1 7 10 0 8

4096 7 0 1 10 8

8192 0 7 10 1 8

16384 9 3 8 2 7

32768 0 10 3 7 6

65536 7 10 5 1 3

131072 5 8 10 1 6

262144 2 10 8 5 6

524288 2 6 4 8 0

1048576 4 1 3 7 9

2097152 1 2 7 5 10

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 4 8 11

4096 6 8 11

8192 8 10 11

16384 9 11 11

32768 2 11 11

65536 11 11 11

131072 11 11 11

262144 3 11 11

524288 11 11 11

1048576 3 11 11

2097152 7 11 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 1.3377e-01 1.3064e-04 31.35 0.05 0.02 0.15

4096 1.2038e-01 2.3512e-04 34.84 0.03 0.01 0.10

8192 1.1334e-01 4.4273e-04 37.01 0.01 0.00 0.06

16384 1.0982e-01 8.5794e-04 38.19 0.01 0.01 0.01

32768 1.0669e-01 1.6670e-03 39.31 0.01 0.01 0.02

65536 1.0689e-01 3.3404e-03 39.24 0.00 0.00 0.00

131072 1.0611e-01 6.6316e-03 39.53 0.00 0.00 0.01

262144 1.0616e-01 1.3269e-02 39.51 0.00 0.01 0.01

524288 1.0504e-01 2.6261e-02 39.93 0.01 0.01 0.02

1048576 1.0347e-01 5.1737e-02 40.54 0.02 0.02 0.04

2097152 1.0319e-01 1.0319e-01 40.64 0.02 0.02 0.03

Five Fastest
Protocols

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

2048 7 1 10 0 8

4096 7 1 10 0 2

8192 10 7 0 1 8

16384 3 2 9 8 7

32768 0 10 2 9 7

65536 9 5 3 4 2

131072 4 10 6 0 9

262144 7 4 2 0 8

524288 1 2 3 10 5

1048576 3 8 7 1 10

2097152 0 9 7 8 1

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 3 8 11

4096 5 8 11

8192 8 9 11

16384 9 11 11

32768 8 11 11

65536 11 11 11

131072 11 11 11

262144 11 11 11

524288 3 11 11

1048576 3 11 11

2097152 3 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 1.3556e-01 1.3238e-04 30.94 0.03 0.02 0.12

4096 1.2073e-01 2.3579e-04 34.74 0.01 0.01 0.04

8192 1.1372e-01 4.4422e-04 36.88 0.01 0.00 0.03

16384 1.1004e-01 8.5970e-04 38.12 0.01 0.00 0.01

32768 1.0691e-01 1.6705e-03 39.23 0.01 0.01 0.02

65536 1.0666e-01 3.3332e-03 39.32 0.01 0.01 0.01

131072 1.0666e-01 6.6665e-03 39.32 0.01 0.01 0.01

262144 1.0684e-01 1.3355e-02 39.26 0.00 0.00 0.01

524288 1.0719e-01 2.6797e-02 39.13 0.01 0.00 0.01

1048576 1.0798e-01 5.3990e-02 38.84 0.03 0.03 0.04

2097152 1.0773e-01 1.0773e-01 38.93 0.01 0.01 0.03

Five Fastest
Protocols

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

2048 1 7 10 0 2

4096 8 2 4 5 9

8192 8 4 2 5 3

16384 3 8 9 2 4

32768 0 10 4 5 3

65536 4 8 9 3 10

131072 3 8 10 7 5

262144 8 2 5 0 6

524288 10 2 8 6 0

1048576 10 8 4 2 6

2097152 3 2 10 8 4

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 2 10 11

4096 10 11 11

8192 10 11 11

16384 6 11 11

32768 5 11 11

65536 10 11 11

131072 11 11 11

262144 11 11 11

524288 8 11 11

1048576 1 11 11

2097152 6 11 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 1.3596e-01 1.3278e-04 30.85 0.02 0.02 0.11

4096 1.2087e-01 2.3607e-04 34.70 0.01 0.00 0.05

8192 1.1394e-01 4.4509e-04 36.81 0.01 0.00 0.02

16384 1.1025e-01 8.6129e-04 38.05 0.01 0.00 0.01

32768 1.0735e-01 1.6774e-03 39.07 0.01 0.01 0.01

65536 1.0683e-01 3.3384e-03 39.26 0.01 0.01 0.01

131072 1.0639e-01 6.6491e-03 39.43 0.01 0.01 0.01

262144 1.0608e-01 1.3260e-02 39.54 0.01 0.01 0.01

524288 1.0670e-01 2.6675e-02 39.31 0.01 0.01 0.02

1048576 1.0920e-01 5.4600e-02 38.41 0.01 0.01 0.02

2097152 1.0760e-01 1.0760e-01 38.98 0.01 0.00 0.03

Five Fastest
Protocols

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

2048 1 7 10 0 3

4096 8 4 2 5 3

8192 8 2 4 3 9

16384 8 9 2 5 3

32768 0 8 2 10 4

65536 2 1 4 8 9

131072 6 4 3 0 10

262144 5 4 8 1 2

524288 4 0 6 8 10

1048576 9 4 6 2 8

2097152 9 3 5 1 7

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 2 10 11

4096 10 10 11

8192 10 11 11

16384 8 11 11

32768 8 11 11

65536 11 11 11

131072 9 11 11

262144 5 11 11

524288 7 11 11

1048576 7 11 11

2097152 6 11 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 1.3612e-01 1.3293e-04 30.81 0.02 0.01 0.11

4096 1.1927e-01 2.3295e-04 35.17 0.01 0.01 0.06

8192 1.1289e-01 4.4096e-04 37.16 0.01 0.01 0.03

16384 1.1033e-01 8.6194e-04 38.02 0.00 0.00 0.01

32768 1.0707e-01 1.6729e-03 39.17 0.01 0.01 0.01

65536 1.0674e-01 3.3356e-03 39.30 0.00 0.00 0.01

131072 1.0623e-01 6.6393e-03 39.48 0.00 0.00 0.01

262144 1.0622e-01 1.3278e-02 39.49 0.01 0.01 0.01

524288 1.0611e-01 2.6528e-02 39.53 0.01 0.01 0.01

1048576 1.0246e-01 5.1231e-02 40.94 0.03 0.04 0.05

2097152 1.0339e-01 1.0339e-01 40.57 0.02 0.01 0.04

Five Fastest
Protocols

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

2048 7 1 8 4 2

4096 4 8 2 9 3

8192 9 3 5 8 4

16384 5 3 9 8 4

32768 10 0 3 2 8

65536 5 2 8 10 4

131072 2 8 6 0 4

262144 10 6 9 2 4

524288 0 10 2 7 4

1048576 10 0 6 7 3

2097152 7 0 3 9 8

Number of Proctocols With
Runtimes Within X% of Min

Msg Size 1% 5% 25%

2048 6 10 11

4096 6 10 11

8192 6 11 11

16384 11 11 11

32768 9 11 11

65536 11 11 11

131072 11 11 11

262144 11 11 11

524288 9 11 11

1048576 1 11 11

2097152 5 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.