CDS 3.2 for GPUs requirements
CDS 3.2 for GPUs has the following requirements:
CDS 3.2 for GPUs requires cluster hosts with NVIDIA Pascal™or better GPUs, with a compute capability rating of 6.0 or higher.
For more information, see Getting Started at the RAPIDS website.
Cloudera and NVIDIA recommend using NVIDIA-certified systems. For more information, see NVIDIA-Certified Systems in the NVIDIA GPU Cloud documentation.
Each cluster host with a GPU must have the following software installed:
- JDK 8 or JDK 11. Cloudera recommends using JDK 8, as most testing has been done with JDK 8. Remove other JDK versions from all cluster and gateway hosts to ensure proper operation.
- Scala 2.12
- Python 3.6 and higher
- GPU drivers and CUDA toolkit
GPU driver v450.80.02 or higher
CUDA version 11.0 or higher
Download and install the CUDA Toolkit for your operating system. The toolkit installer also provides the option to install the GPU driver.
- (Optional) UCX
Clusters with Infiniband or RoCE networking can leverage Unified Communication X (UCX) to enable the RAPIDS Shuffle Manager. For information on UCX native libraries support, see (Optional) Installing UCX native libraries.
Supported versions of CDP are described below.
|CDS 3.2 for GPUs||Supported CDP Versions|
|220.127.116.11.2.7170.0-49||CDP Private Cloud Base with Cloudera Runtime 7.1.7 and higher|
The Spark 2 service (included in CDP) can co-exist on the same cluster as a Spark 3 service (including CDS 3.2 for GPUs, installed as a separate parcel). The two services are configured to not conflict, and both run on the same YARN service. CDS 3.2 for GPUs installs and uses its own external shuffle service.
Although Spark 2 and Spark 3 can coexist in the same CDP Private Cloud Base cluster, you cannot use multiple Spark 3 versions simultaneously. All clusters managed by the same Cloudera Manager Server must use exactly the same version of CDS, whether that is CDS Powered by Apache Spark or CDS 3.2 for GPUs.