High-Performance Computing-Center Stuttgart, HLRS: Eff Eff Eff

You might also like

You are on page 1of 5

Eective Bandwith Benchmark (bef f ) Version 3.

5
High-Performance Computing-Center Stuttgart, HLRS

Thu Jan 24 13:21:20 2013 on CNK r00idj07 2.6.32-220.23.3.bgq.el6 V1R1M2 0.ppc64 1 BGQ be = 89490.849 MB/s = 43.697 * 2048 PEs with 128 MB/PE
number of processors bef f Lmax bef f at Lmax rings& random MByte/s 191847 204 bef f at Lmax rings only MByte/s 417199 Latency rings& random mircosec 8.585 Latency rings only microsec 8.056 Latency pingpong microsec 4.882 ping-pong bandwidth

accumulated per process

2048

MByte/s 89491 44

1 MB 94

MByte/s 2211

Ping-Pong result (only the processes with rank 0 and 1 in MPI COMM WORLD were used): Latency: 4.882 microsec per message Bandwidth: 2210.760 MB/s (with MB/s = 106 byte/s)

Sndrcv, ring & random patterns 10000

1000

bandwith [MB/s]

100

10 ring-1024*2fix ring-512*4fix ring-256*8fix ring-4*512fix ring-2*1024fix ring-1*2048fix worst random avg random best random 1 10 100 1000 10000 message length per process [Byte] Sndrcv, additional patterns 10000 100000 1e+06

0.1

0.01

1000

bandwith [MB/s]

100

10 worst-cyc-1dim best bi-section worst bi-section acyclic-2dim-all acyclic-3dim-all cyclic-2dim-x cyclic-2dim-y cyclic-2dim-all cyclic-3dim-x 1 10 100 1000 10000 message length per process [Byte] 100000 1e+06

0.1

0.01

Alltoal, ring & random patterns 10000

1000

bandwith [MB/s]

100

10 ring-1024*2fix ring-512*4fix ring-256*8fix ring-4*512fix ring-2*1024fix ring-1*2048fix worst random avg random best random 1 10 100 1000 10000 message length per process [Byte] Alltoal, additional patterns 10000 100000 1e+06

0.1

0.01

1000

bandwith [MB/s]

100

10 worst-cyc-1dim best bi-section worst bi-section acyclic-2dim-all acyclic-3dim-all cyclic-2dim-x cyclic-2dim-y cyclic-2dim-all cyclic-3dim-x 1 10 100 1000 10000 message length per process [Byte] 100000 1e+06

0.1

0.01

non-blk, ring & random patterns 10000

1000

bandwith [MB/s]

100

10 ring-1024*2fix ring-512*4fix ring-256*8fix ring-4*512fix ring-2*1024fix ring-1*2048fix worst random avg random best random 1 10 100 1000 10000 message length per process [Byte] non-blk, additional patterns 10000 100000 1e+06

0.1

0.01

1000

bandwith [MB/s]

100

10 worst-cyc-1dim best bi-section worst bi-section acyclic-2dim-all acyclic-3dim-all cyclic-2dim-x cyclic-2dim-y cyclic-2dim-all cyclic-3dim-x 1 10 100 1000 10000 message length per process [Byte] 100000 1e+06

0.1

0.01

Best transfer method, ring & random patterns 10000

1000

bandwith [MB/s]

100

10 ring-1024*2fix ring-512*4fix ring-256*8fix ring-4*512fix ring-2*1024fix ring-1*2048fix worst random avg random best random 1 10 100 1000 10000 100000 message length per process [Byte] Best transfer method, additional patterns 10000 1e+06 1e+07

0.1

0.01

1000

bandwith [MB/s]

100

10 worst-cyc-1dim best bi-section worst bi-section acyclic-2dim-all acyclic-3dim-all cyclic-2dim-x cyclic-2dim-y cyclic-2dim-all cyclic-3dim-x 1 10 100 1000 10000 100000 message length per process [Byte] 1e+06 1e+07

0.1

0.01

Ring & random average: Sndrcv, Alltoal, non-blk 10000

1000

bandwith [MB/s]

100

10

1 Sendrcv rings Alltoal rings non-blk rings Sendrcv random Alltoal random non-blk random 1 10 100 1000 10000 message length per process [Byte] Best method: rings & random 10000 100000 1e+06

0.1

0.01

1000

bandwith [MB/s]

100

10

0.1

0.01 1 10

rings minumum rings average rings maximum random minimum random average random maximum ring & random average 100 1000 10000 100000 message length per process [Byte] 1e+06 1e+07

You might also like