HOME PAGE FOR VMI21

[NCSA]
Virtual Machine Interface 2.1

VMI and MPICH-VMI Latency


The VMI latency data graphed below was measured using the latency benchmark installed in the benchmarks directory of VMI install tree. latency measures the ping-pong latency of an interconnect by averaging the latency for n messages, each of size s, where n and s is specified by the user.

The MPICH-VMI latency was measured using latency_mpi benchmark installed in the tools directory of MPICH-VMI install tree. latency_mpi calculates both ping-pong latency (round-trip latency) and one-way latency of the interconnect using the MPI messaging interface.

MPICH-VMI, by default, switches from eager protocol (implemented as streams) to rendezvous protocol (implemented as RDMAs) for sending/receiving data of size 16k and over (The default message length over which rendezvous is used can also be set by the user on command line). So for all MPICH-VMI latency curves plotted below, data points over 16k message size are measurements of latency for rendezvous protocol over RDMA.

IA-32 Latency Measurements

To meaure bandwidth between IA-32 machines, we used 2 Dell PowerEdge 1750 servers with 2 hyperthreaded Intel Xeon 3.06 GHz processors (3 GB RAM, 512k L2 cache) running Linux 2.4.20-30smp kernels for all bandwidth measurements.

For Infiniband, both machines had InfiniHost MT23108 HCA (a1 silicon) on 64 bit PCIX 133MHz bus with thca-3.1.

Msg Size (Bytes) VMI Latency (usec) MPI Latency (usec) MPI Latency (usec)
0 1.77 5.35
1 7.4815 1.74 5.34
2 7.7615 1.78 5.34
4 7.539 1.74 5.35
8 7.6405 1.75 5.17
16 7.701 1.78 5.27
32 7.7645 1.77 6.23
64 8.935 1.86 6.38
128 9.172 1.85 6.63
256 9.761 1.91 7.14
512 10.943 1.94 8.13
1024 13.26 2.17 10.11
2048 14.09 2.79 12.01
4096 16.311 3.92 15.91
8192 21.204 6.38 23.57
16384 27.599 21.99 41.49
32768 45.91 31.35 60.12
65536 82.534 50.32 97.83
131072 155.91 88.08 173.2
262144 302.69 163.67 324.14
524288 767.82 315.02 626.03
1048576 1580.9 617.7 1229.82
2097152 3157.5 1230.85 2464.59
4194304 6369.4 2449.79 4912.63

For Myrinet, both machines had LANai 10 on 64 bit PCIX 100MHz bus with gm-2.0.11 driver.

Msg Size (Bytes) VMI Latency (usec) MPI Latency (usec) MPI Latency (usec)
0 3.06 8.1
1 6.559 3.04 8.29
2 6.5475 3.04 8.17
4 6.5505 3.05 8.16
8 6.6135 3.04 7.87
16 6.8275 3.05 8.06
32 6.7745 3.05 8.22
64 6.9785 3.07 8.38
128 8.0425 3.1 8.91
256 10.412 3.13 9.8
512 12.775 3.28 11.41
1024 16.117 3.52 14.86
2048 22.362 4.38 21.97
4096 31.069 8.5 31.78
8192 49.44 16.8 52.12
16384 84.442 51.25 99.88
32768 150.83 84.57 166.1
65536 283.11 150.92 298.65
131072 548.13 283.57 563.51
262144 1079.8 548.74 1096.66
524288 2143.2 1079.28 2153.68
1048576 4259.5 2151.03 4273.87
2097152 8498.9 4273.02 8526.98
4194304 16980 8510.22 17010.5

for Ethernet, Broadcom Gigabit Ethernet NIC with bcm5700 driver was used.

Msg Size (Bytes) VMI Latency (usec) MPI Latency (usec) MPI Latency (usec)
0 67.29 61.7
1 63.876 12.52 59.6
2 63.825 12.53 59.53
4 63.648 12.59 59.51
8 63.585 12.52 59.65
16 64.417 12.55 60.01
32 66.872 12.64 60.25
64 67.19 12.66 61.99
128 69.098 12.84 63.78
256 73.893 13.11 67.84
512 81.436 12.88 76.17
1024 95.627 13.72 90.2
2048 118.56 13.79 115.93
4096 171.4 21.67 166.75
8192 281.16 38.61 282.81
16384 346.35 203.39 783.67
32768 473.5 267.55 608.13
65536 748.15 407.03 979.32
131072 1279.8 670.2 1398.55
262144 2369 1238.92 3102.5
524288 4487.2 2280.46 4634.07
1048576 8767.1 4422.39 9152.83
2097152 17282 8689.43 17438.63
4194304 34325 17201.88 34447.29

IA-64 Latency Measurements


To meaure bandwidth between IA-64 machines over Myrinet, we used 2 Intel Itanium 2 duals with 1.3 Ghz processors (4 GB RAM) running Suse SLES8 system with 2.4.21 kernel. Infiniband, GigE and shared memory measurement were done with 2 Intel Itanium 2 duals with 900 Mhz processors (2 GB RAM) running 2.4.21 kernel.

For Infiniband, both machines had InfiniHost MT23108 HCA (a1 silicon) on 64 bit PCIX 133MHz bus with thca-3.1.
Msg Size (Bytes) VMI Latency (usec) MPI Latency (usec) MPI Latency (usec)
0 3.96 7.97
1 10.193 3.98 7.97
2 10.17 3.96 7.95
4 10.213 3.97 7.95
8 10.243 3.97 8.03
16 10.301 3.88 8.77
32 10.487 3.84 8.8
64 11.694 3.92 9.05
128 11.901 3.91 9.42
256 12.233 3.94 10.05
512 12.857 3.94 11.35
1024 14.713 4.03 13.27
2048 15.831 4.15 15.4
4096 17.858 4.8 19.78
8192 23.048 6.78 28.55
16384 33.666 28.61 54.43
32768 55.284 39.64 76.47
65536 98.578 61.74 120.72
131072 185.03 106.03 209.28
262144 357.93 194.65 386.32
524288 703.7 371.92 740.63
1048576 1394.9 726.27 1448.75
2097152 2780.5 1439.38 2881.57
4194304 5772.1 2869.08 5752.88

For Myrinet, both machines had LANai 10.0 on 64 bit PCIX 100MHz bus with gm-2.0.11 driver.
Msg Size (Bytes) VMI Latency (usec) MPI Latency (usec) MPI Latency (usec)
0 3.74 11.93
1 9.8315 3.63 12.84
2 11.19 3.61 12.1
4 10.059 3.57 12.44
8 10.958 3.55 12.66
16 10.987 3.56 13.07
32 11.787 4.08 13.04
64 12.235 3.62 12.58
128 14.027 3.78 12.87
256 16.507 4.09 14.82
512 18.097 3.91 17.31
1024 21.661 4.37 20.8
2048 29.29 5.77 28.49
4096 41.239 9.39 42.3
8192 66.527 17.66 66.96
16384 98.584 62.19 118.49
32768 167.79 102.26 190.25
65536 311.48 177.99 334.06
131072 604.75 316.03 625.26
262144 1175.5 621.27 1210.48
524288 2340 1214.98 2370.88
1048576 4677.4 2395.16 4690.53
2097152 9316.2 4753.94 9343.18
4194304 18612 9468.52 18654.69

Broadcom Gigabit Ethernet NIC with bcm5700 driver was used for Ethernet.

Msg Size (Bytes) VMI Latency (usec) MPI Latency (usec) MPI Latency (usec)
0 16.94 73.68
1 81.483 15.65 72.63
2 83.221 16.55 73.78
4 83.126 15.28 72.5
8 81.521 16.76 73.94
16 83.673 15.18 72.9
32 81.658 16.22 74.95
64 82.608 15.05 73.92
128 84.894 15.71 77.98
256 89.421 17.43 80.11
512 96.609 15.67 89.7
1024 111.87 16.97 103.39
2048 130.96 19.27 122.44
4096 155.5 23.69 147.89
8192 176.8 35.77 172.17
16384 254.44 171.26 407.17
32768 394.12 236.05 541.55
65536 706.22 386.82 928.86
131072 1276.1 720.57 1562.82
262144 2464.4 1256.09 2788.56
524288 4629 2446.51 4859.01
1048576 9102.6 4631.63 9345.74
2097152 18126 9108.6 18355.93
4194304 36034 18111.74 36304.89

Shared memory.

Msg Size (Bytes) MPI Latency (usec) MPI Latency (usec)
0 0.46 1.66
1 0.53 1.78
2 0.47 1.72
4 0.48 1.74
8 0.5 1.72
16 0.48 1.73
32 0.51 1.77
64 0.45 2.08
128 0.5 2.18
256 0.53 2.44
512 0.6 2.91
1024 0.77 3.43
2048 1.13 4.33
4096 1.91 6.1
8192 3.48 9.68
16384 9.32 19.55
32768 16.12 33.32
65536 29.46 60.3
131072 55.9 114.07
262144 108.33 222.47
524288 224.28 458.8
1048576 710.98 1508.33
2097152 1902.46 3642.25
4194304 3891.65 7575.39

Back to the VMI2 web page.

 


[NCSA]