Administrator Guide

7 RAPIDS Scaling on Dell EMC PowerEdge Servers
Figure 3. Average ranking of the ML algorithms.
Source: Nvidia/ https://arxiv.org/pdf/1708.05070.pdf
1.2 Dask for Parallel Computing
Dask is a distributed computation scheduler built to scale Python workloads from laptops to supercomputer
clusters. It takes advantage of hardware advancements using a communications framework called
OpenUCX to scale up and out with cuDF (Native integration with Dask + cuDF):
• For intranode data movement, utilizing NVLink and PCIe peer-to-peer communications
• For internode data movement, utilizing GPU RDMA over InfiniBand and RoCE
Figure 4. Dask for Parallel Computing. Source: Nvidia