The model training runtime improvement for each data science pipeline used in this test with Granulate (sAgent enabled) compared with baseline (sAgents deactivated) is shown in Figure 3. The results show that an improvement of at least 10% was always obtained. The average runtime improvement across all data science pipelines is 23%, while the maximum improvement was 37% for the LogR data science pipeline.

All the improvements were obtained by enabling the sAgents in the systems. No application code changes were made, and no fine-tuning of the platform or operating system settings was required.

Figure 3.   Training time improvement with Granulate compared to baseline

A separate test was run to optimize cluster CPU utilization metrics during model training followed by serving of the various data science pipelines. Granulate’s sAgent reduced CPU utilization up to 40% for the Sales forecasting pipeline. The lowest improvement on CPU utilization was 11% for Customer segmentation. See Figure 4 for the various data science pipelines that were optimized using Granulate.

In a production environment, this optimization would result in:

• Reducing the rack space needed in the data center, thus reducing the Total Cost of Ownership (TCO)
• Scaling up the workloads—for instance, by increasing the dataset sizes to use the freed-up compute resources

Figure 4.   CPU Utilization reduction for different data science pipelines

Table 2.      System under test characteristics