Performance
Test environment
All performance tests decribed in the following are run on the Kepler Cluster. Within the Kepler Cluster the nodes having a Dual AMD Athlon MP 2000+ processor with 1666 MHz, 256 L2 Cache and 1-2GB RAM are used. Each node is running Linux (2.4.21 kernel) and communication occurs over a Myrinet interconnect with MPI Peak performance of 115MB per second and node. Due to technical reasons it is not possible to run πSvM on more than eight processors. Therefore the following performance figures show speedup and CPU time on up to eight processors.
Datasets
The datasets are available for download here. For convenience here is a list of the original data sources:Classification Performance
To test parallel classification performance the four datasets listed in the following table are used. The datasets are available for download here.
| Dataset | Train Size | Test Size | Dimension |
|---|---|---|---|
| mnist-576-rbf-8vr | 60000 | 10000 | 576 |
| mnist-784-poly-8vr | 60000 | 10000 | 784 |
| covtype-2vr | 435759 | 145253 | 54 |
| kddcup99-nvr | 4898430 | 311029 | 122 |
For C-SVC and ν-SVC training πSvM achieves superlinear speedups. Slight deviations from this behavior are the runtime of C-SVC on the kddcup-99-nvr dataset and the runtime of ν-SVC on the mnist-576-rbf-8vr dataset. A comparison with respect to single processor performance of πSvM and LibSVM can be found here.
Performance of C-SVC
C-SVC Speedup and CPU time for dataset mnist-576-rbf-8vr.
C-SVC Speedup and CPU time for dataset mnist-784-poly-8vr.
C-SVC Speedup and CPU time for dataset covtype-2vr.
C-SVC Speedup and CPU time for dataset kddcup99-nvr.
Performance of ν-SVC
ν-SVC Speedup and CPU time for dataset mnist-576-rbf-8vr.
ν-SVC Speedup and CPU time for dataset mnist-784-poly-8vr.
ν-SVC Speedup and CPU time for dataset covtype-2vr.
ν-SVC Speedup and CPU time for dataset kddcup99-nvr.
Regression Performance
To test parallel regression performance the two datasets listed in the following table are used. The datasets are available for download here.
| Dataset | Train Size | Test Size | Dimension |
|---|---|---|---|
| kddcup98 | 95412 | 96367 | 403 |
| mv | 36768 | 4000 | 10 |
In the regression setting the πSvM speedup is close to linear only for the kddcup98 dataset for both ε-SVR and ν-SVR. For the mv dataset a linear speedup with πSvM is only possible when training ε-SVR on two processors. The reasons for this breakdown of performance are still under invesitgation but it is assumed that the speedup potential for this dataset is low since the single processor runtime is already quite short (20min). A comparison with respect to single processor performance of πSvM and LibSVM can be found here.
Performance of ε-SVR
ε-SVR Speedup and CPU time for dataset kddcup98.
ε-SVR Speedup and CPU time for dataset mv.
Performance of ν-SVR
ν-SVR Speedup and CPU time for dataset kddcup98.
ν-SVR Speedup and CPU time for dataset mv.