NUMA Effects

Latency (Sequential)

Latency Per Batch

ResNet Latency BS=1 ResNet Latency BS=2 ResNet Latency BS=4 ResNet Latency BS=8 ResNet Latency BS=16 ResNet Latency BS=32 ResNet Latency BS=64 ResNet Latency BS=128 ResNet Latency BS=256 ResNet Latency BS=512 ResNet Latency BS=1024

Latency Per Optimization

ResNet Latency Opt=Script

Throughput (Sequential)

Throughput Per Batch

ResNet Throughput BS=1 ResNet Throughput BS=2 ResNet Throughput BS=4 ResNet Throughput BS=8 ResNet Throughput BS=16 ResNet Throughput BS=32 ResNet Throughput BS=64 ResNet Throughput BS=128 ResNet Throughput BS=256 ResNet Throughput BS=512 ResNet Throughput BS=1024

Throughput Per Optimization

ResNet Latency Opt=Script

Speedup

ResNet Latency Opt=Script

Throughput Per Batch

ResNet Throughput per BS

NUMA-Effects

% Latency Degradation HeatMap

ResNet Latency Degradation

% Throughput Degradation HeatMap

ResNet Latency Degradation

Latency Per Batch

ResNet Latency BS=1 ResNet Latency BS=2 ResNet Latency BS=4 ResNet Latency BS=8 ResNet Latency BS=16 ResNet Latency BS=32 ResNet Latency BS=64 ResNet Latency BS=128 ResNet Latency BS=256 ResNet Latency BS=512 ResNet Latency BS=1024

% Latency Loss Per Batch

ResNet Latency BS=1 ResNet Latency BS=2 ResNet Latency BS=4 ResNet Latency BS=8 ResNet Latency BS=16 ResNet Latency BS=32 ResNet Latency BS=64 ResNet Latency BS=128 ResNet Latency BS=256 ResNet Latency BS=512 ResNet Latency BS=1024

Throughput Per Batch

ResNet Throughput BS=1 ResNet Throughput BS=2 ResNet Throughput BS=4 ResNet Throughput BS=8 ResNet Throughput BS=16 ResNet Throughput BS=32 ResNet Throughput BS=64 ResNet Throughput BS=128 ResNet Throughput BS=256 ResNet Throughput BS=512 ResNet Throughput BS=1024

% Throughput Loss Per Batch

ResNet Throughput BS=1 ResNet Throughput BS=2 ResNet Throughput BS=4 ResNet Throughput BS=8 ResNet Throughput BS=16 ResNet Throughput BS=32 ResNet Throughput BS=64 ResNet Throughput BS=128 ResNet Throughput BS=256 ResNet Throughput BS=512 ResNet Throughput BS=1024