Configuring a ML for IBM z/OS cluster for high availability
If your machine learning workload is large, mission critical, or both, you can configure your MLz core services for high availability. To achieve high availability of core services, consider configuring a MLz cluster. Each cluster can consist of two or more MLz instances that run either on a single LPAR or across different LPARs.
Before you begin
A MLz instance contains a set of core services for model training, deployment, batch scoring, ingestion, repository, and data connection management, and the core services are supported by an active runtime environment. The runtime environment can be provided by Spark, Python, or both. To make core services in a MLz cluster highly available means to keep one runtime environment active at all time.
- Decide the type of MLz cluster
you want to configure. As shown in the following diagram, you can configure a cluster with multiple
MLz instances running on the same
LPAR (Cluster type 1) or across different LPARs (Cluster type 2).
Figure 1. MLz cluster
- Set up the TCP SHAREPORT port or the sysplex distributor port to be used by the cluster. If your MLz cluster is type 1, enable the SHAREPORT. If your cluster is type 2, enable the sysplex distributor port.
- Provision and install additional system capacity to support your MLz cluster.
It is recommended that you plan and start each MLz instance in a cluster, regardless of the cluster type, with the basic system capacity as described in Planning system capacity for ML for IBM z/OS. You can adjust the basic capacity in terms of CPU, memory, or DASD over time based on your machine learning workload.
If your cluster is type 1 and if your initial workload is small, you might be able to share the basic capacity of 1 GCP, 4 zIIPs, and 100 GB memory across multiple MLz instances on the same LPAR because only one runtime environment is active at any given time. You can increase the CPU and memory allocation as your workload increases. However, you must allocate and start with 100 GB DASD for each instance.