metaflow_sysml19

Repository for SysML19 Artifacts Evaluation

What have done

Test the latency of two demo networks in the original paper

SampleA:

SampleB:

The channel of input is 128, image size ranges in 28x28, 56x56, 112x112 (image size of the block in ResNet should be 28x28)

The batchsize ranges in 1, 8

The network definition and latency testing code can be found in code/cnn.cc

Results on Tesla V100

Outputs:

Batchsize 1  image size 28x28
SampleA: 0.19661 0.19661 0.19681 0.19651 0.19702 0.19630
Mean Latency 0.19664  Std 0.00056
SampleB: 0.27863 0.27873 0.27894 0.27894 0.28006 0.28027
Mean Latency 0.27926  Std 0.00160

Batchsize 8  image size 28x28
SampleA: 0.31212 0.31201 0.31181 0.31181 0.31171 0.31150
Mean Latency 0.31183  Std 0.00049
SampleB: 0.47657 0.47657 0.47667 0.47636 0.47636 0.47677
Mean Latency 0.47655  Std 0.00037

Batchsize 1  image size 56x56
SampleA: 0.21504 0.21504 0.21494 0.21484 0.21473 0.21555
Mean Latency 0.21502  Std 0.00064
SampleB: 0.27597 0.27587 0.27587 0.27576 0.27628 0.27637
Mean Latency 0.27602  Std 0.00055

Batchsize 8  image size 56x56
SampleA: 0.89487 0.89487 0.89508 0.89508 0.89477 0.89477
Mean Latency 0.89491  Std 0.00031
SampleB: 1.38455 1.38261 1.38179 1.39233 1.39315 1.37411
Mean Latency 1.38476  Std 0.01596

Batchsize 1  image size 112x112
SampleA: 0.51210 0.51210 0.51200 0.51200 0.51190 0.51190
Mean Latency 0.51200  Std 0.00020
SampleB: 0.69786 0.69785 0.69775 0.69775 0.69806 0.69806
Mean Latency 0.69789  Std 0.00031

Batchsize 8  image size 112x112
SampleA: 3.00534 3.00616 3.00360 3.00329 3.00278 3.00236
Mean Latency 3.00392  Std 0.00335
SampleB: 4.74286 4.74317 4.74255 4.74245 4.74388 4.74429
Mean Latency 4.74320  Std 0.00166

System information:

• CPU: Intel Xeon E5-2690v4

• GPU: Tesla V100-PCIE

• Nvidia Driver: 410.104

• CUDA version: 10.0.130

• CUDNN version: 7.6.2.24

Reproduce:

cd code; make gpu; cd ..; ./mf