Benchmarks

Benchmark for:

• xgboost speed
• LightGBM speed

To compare:

• Non-Uniform Memory Access (NUMA) and Uniform Memory Access (UMA)
• Model (xgboost exact, approx, hist, LightGBM)

Dataset used:

• Bosch with less than 95% sparse features

Server used:

• Dual Xeon Gold 6154 (36 cores / 72 threads, 3.7/3.7 GHz)

Benchmark for:

• xgboost speed
• LightGBM speed

To compare:

• Non-Uniform Memory Access (NUMA) and Uniform Memory Access (UMA)

Dataset used:

• Feature Interaction Higgs (11M x 406) => 82GB RAM required for benchmarking

Server used:

• Dual Xeon Gold 6130 (32 cores / 64 threads, 3.7/2.8 GHz)

Benchmark for:

• xgboost speed
• LightGBM speed

To compare:

• gcc 5.4, gcc 8.1, and Intel Compilers 2018.2.199 (from Intel Parallel Studio XE 2018 Update 2)

Dataset used:

• Feature Interaction Higgs (11M x 406) => 82GB RAM required for benchmarking

Server used:

• Dual Xeon Gold 6130 (32 cores / 64 threads, 3.7/2.8 GHz)

Benchmark for:

• xgboost speed
• LightGBM speed

To compare:

• Windows Subsystem for Linux (WSL) and Windows

Dataset used:

• Feature Interaction Higgs (11M x 406) => 82GB RAM required for benchmarking

Server used:

• Dual Xeon Gold 6130 (32 cores / 64 threads, 3.7/2.8 GHz)

Benchmark for:

• xgboost
• LightGBM

To compare:

• Speed and accuracy of models

Datasets used:

• Bosch (1M x 1K)
• Higgs (10M x 29)
• Reputation (2M x 25K)

Servers used:

• Intel Core i7-7700K (4 cores / 8 threads, 4.5/4.4 GHz)
• Dual Xeon E5-2670v2 (20 cores / 40 threads, 3.3/2.7 GHz)

Benchmark for:

• xgboost
• LightGBM

To compare:

• Speed and accuracy of models

Datasets used:

• Bosch (1M x 1K)
• Higgs (10M x 29)
• Reputation (2M x 25K)

Servers used:

• Intel Core i7-7700K (4 cores / 8 threads, 4.5/4.4 GHz)
• Dual Xeon E5-2670v2 (20 cores / 40 threads, 3.3/2.7 GHz)