On-Premises Deployment Specifications

On This Page

Overview

This document lists the hardware specifications for on-premises ("on-prem") deployment of version 3.5.5 of the MLOps Platform ("the platform").

Capacity Calculations
All capacity calculations in the hardware specifications are performed using the base-10 (decimal) number system. For example, 1 TB = 1,000,000,000,000 bytes.
Note
The same hypervisor can be used for both data and application nodes.
Docker Registries
The platform is deployed with a default, on-cluster docker registry. This registry can only be used for playground environments, as it is volatile and images are lost when it is restarted. Operational clusters must be connected to an external Docker registry. Create the registry in your cluster, then configure it post-installation through the platform dashboard.

Hardware Configurations

Iguazio Playground

A single data-node and single application-node cluster implementation. This configuration is designed mainly for evaluation trials and doesn't include high availability (HA) or performance testing.

Note
You can deploy a Proof of Concept (POC) in the Iguazio playground configuration. Be aware that the POC cannot be used as development environment or for a production environment.

For both configurations, data nodes in on-prem deployments are always deployed on virtual machines (VMs) while application nodes can be deployed on either VMs or local machines (bare-metal).

GCP Production Cluster

VM Deployment Notes

Warning

  • Provisioning (deployment) of the platform's node VMs is done by using a dedicated virtualization package, provided by Iguazio.
    Don't attempt to provision the servers yourself prior to the deployment.

  • Platform shutdown should be coordinated with Iguazio's support team.
    Don't shut down the data-node VMs non gracefully, as this might erase the data.

When deploying on virtual machines, notify Iguazio's support team whenever VMware Enhanced vMotion Compatibility (EVC) mode is enabled, as a low EVC level might disable required CPU features.

VM Hypervisor Host Specifications

Hypervisor host machines in VM platform deployments must fulfill the following hardware specification requirements:

Component Specification
Network interfaces
  • A single-port 1 Gb (minimum) NIC for the management network
  • For hosting data-node VMs — a dual-port 10 Gb (minimum) NIC for the data-path (client) and interconnect networks
  • For hosting application-node VMs only — a single-port 10 Gb (minimum) NIC for the data-path (client) network
Hypervisor VMware vSphere ESXi 6.5, 6.7, 7.0, or Proxmox VE (PVE) 6.x or 7.0
Note

If you want to configure your VM Hypervisor for High Availability:

  • Each of the VM Data Nodes must run on different hypervisors.
  • For at least 3 application nodes: each of those VM Application Nodes must run on different hypervisors. The rest of the VM Application Nodes can run more than one on same hypervisors.
  • One VM Application node and one VM Data node can use same hypervisor

VM Data-Node Specifications

Data nodes in on-prem platform deployments are VMs that must fulfill the following hardware specification requirements.

Note

  • For some components, the specification differentiates between small and large data nodes. Large data nodes provide greater processing capabilities.
    You cannot mix the specifications of these two alternative configurations.
  • CPU Hyper-threading is not supported in the data cluster.
Component Specification
Memory 64 GB (POC node) / 128 GB (small node) / 256 GB (large node)
Cores 4 (POC node) / 8 (small node) / 16 (large node)
VM boot disk 400 GB (minimum) image (hosted on an enterprise-grade SSD-based data store)
Data disks 2 (POC node) / 2, 4, 6, or 12 (small node) / 2, 4, 6, 8, or 12 (large node) NVMe or enterprise-grade SSD data disks (drives) of 1 TB (minimum) each, which are mapped exclusively to the data-node VM using direct attach storage (DAS), such as raw device mapping (RDM).

Application-Node Specifications

In on-prem deployments you can select whether to deploy the application nodes on virtual machines (VMs) or on local machines (bare-metal), provided the same method is used on all nodes.

VM Application-Node Specifications

Application nodes in VM platform deployments are VMs that must fulfill the following hardware specification requirements.

Note
For some components, the specification differentiates between small and large application nodes. Large application nodes provide greater processing capabilities. Note that you cannot mix the specifications of these two alternative configurations.
Component Specification
Memory 64 GB (small node) / 128 GB (large node)
Cores 8 (small node) / 16 (large node)
VM boot disk 400 GB (minimum) image (hosted on an enterprise-grade SSD-based data store)
GPU (Optional) GPUs supported by the NVIDIA Linux display driver 470.129.

Bare-Metal Application-Node Specifications

Application nodes in bare-metal platform deployments are supplied by the customer and must fulfill the following hardware specification requirements:

Cores 8 (minimum)
Memory 64 GB of RAM (minimum)
OS boot disk 400 GB (minimum) enterprise-grade SSDs
Network interfaces
  • Single port 1 Gb (minimum) NIC for the management network
  • Single port 10 Gb (minimum) NIC for the data-path (client) network
GPU (Optional) GPUs supported by the NVIDIA Linux display driver 470.129.

Backing up the Platform

Caution
To ensure safety of your MLOps Platform data, you must periodically back up your data and configuration.

Allocate one NFS in your application node for backing up your platform. See full details in Backing Up the Platform.

See Also