Skip to content

Latest commit

 

History

History
111 lines (82 loc) · 6.68 KB

deployment-planning.md

File metadata and controls

111 lines (82 loc) · 6.68 KB
Raw

Planning a Graphistry Deployment

You can architect your deployment based on usage mode and environment constraints. We recommend starting with AWS/Azure Marketplace for most projects.

The following documents describe common Graphistry configurations, examples of which configuration may make sense for you, Graphistry's components, and planning for multi-year and RAPIDS scaleout.

Related:

1. Common Configurations

Graphistry deployments generally follow one of these patterns:

  1. Preconfigured AWS/Azure Marketplace instance
    The most typical case.
  • Preconfigured for one-click quicklaunch: Nvidia, Docker, Graphistry, Jupyter, connectors, and many DB/API samples
  • Data stays in your team's private cloud account
  • Great for POCs. Utility-fee pricing means cheap POCs by turning off the instance when inactive.
  • Works well with databases also running securely in your cloud account and external APIs
  • Dedicated Graphistry support
  1. Preconfigured AWS/Azure Marketplace instance with on-prem data bridging
    On-prem enterprise data sources can likely be accessed as-is or via firewall changes. However, sometimes it is easier to use an on-prem jump box.
  • Start with Graphistry AWS/Azure Marketplace as usual, e.g., for POCs
  • Run the Graphistry Data Bridge closer to the DB/API, e.g., on-prem
  • Configure the Graphistry Data Bridge, Graphistry server, and firewalls/DBs to work in bridged mode
  1. Manual GPU environment with Graphistry containers
    Primarily useful for heavily regulated environments such as Federal and Financial teams with requirements such as air-gapping.
  • Start with Graphistry AWS/Azure Marketplace as usual, e.g., for POCs
  • When ready, follow Graphistry guides for Nvidia/Docker environment setup and Graphistry container installation
  • Likely benefits from online Graphistry Support
  • We recommend Cloud over On-Prem when possible
  1. Graphistry Cloud (Alpha)
    Graphistry, Inc. runs and manages your Graphistry account, with a choice of shared and isolated tiers:
  • Isolated tier: Useful when it is difficult to one-click launch on the Marketplace or procure/setup on-prem hardware
  • Shared tier: Useful if you want the cost savings and expect limited use
  1. Virtual machine Graphistry can run in a virtual machine (VMWare, Nutanix, ...) that shares its GPU(s) with other virtual machines:
  • Graphistry will still run as a Docker container in the VM
  • Depending on the hypervisor, you can pick different vGPU sharing modes, and further partition within the VM via Docker settings
  • vGPUs do not currently support CUDA Unified Memory, so you must configure Graphistry setting RMM_ALLOCATOR=default in data/config/custom.env
  • See our more detailed vGPU docs

2. Picking a Configuration

  • POC:
    Save significant time by going with AWS/Azure Marketplace!

    • Optional: Turn off Graphistry when not using it
    • Do not use if you cannot put your data in the cloud
    • Dedicated Graphistry staff is happy to assist with setup, configuration, demos, use cases, and other POC tasks
    • If needed, Graphistry staff can help with Data Bridge setup

  • Analyst team:
    Go with AWS/Azure Marketplace or AWS/Azure BYOL

    • Do not use if you cannot put your data in the cloud
    • Dedicated Graphistry staff is happy to assist with setup, configuration, demos, use cases, and other POC tasks
    • If needed, Graphistry staff can help with Data Bridge setup

  • Analyst team, no AWS/Azure/GPU access:
    Go with Graphistry Cloud (alpha). Please contact Graphistry for discussion of options and support.

  • Highly-regulated federal, bank, telco, utility, hospital:
    Get started with Graphistry Marketplace for the POC, and go on-prem from there.

  • Commercial Product Developer:
    To empower your product with Graphistry, we recommend POC'ing via Marketplace and using Graphistry Cloud's API (alpha) similar to how you might use Google Maps.

3. Graphistry Components

Understanding Graphistry's architecture may help you make your deployment architecture decisions:

  • Graphistry Core: The visualization server
    • GPU Visualization uploads and live sessions
    • Docker containers that require nvidia-docker and Nvidial Pascal or later (See GPU/OS hardware requirements)
    • Supports multi-GPU (single node) and GPU resource isolation
    • Configurable reverse proxy, TLS, and user/API key management
  • Graphistry Investigation Server (Optional)
    • Installed as part of Graphistry Core
    • Investigation templates
    • Investigation cases, with embedded visualizations/sessions
    • Connectors: Pushdown queries over DB/API connectors
    • Deep linkable: External systems (Splunk, dashboards, emails, ...) can be configured with links into Graphistry views and workflows
  • Graphistry Data Bridge (Optional)
    • Standalone docker-based server (See Graphistry Data Bridge)
    • Enables individual Investigation Server connectors to proxy through a jump server
    • Useful for bridging Graphistry cloud with on-prem data sources when simpler alternatives do not suffice
  • Graphistry API
    • Used by external web apps, notebook servers, etc.
    • Communicates via language-neutral REST APIs and Graphistry-provided per-language convenience libraries
  • Data Science Notebooks
    • Jupyter is installed as part of Graphistry Core
    • External notebooks may also be used via Graphistry API

4. Planning for Multi-Year and RAPIDS Scaleout

Graphistry helped start and integrates the growing RAPIDS GPU computing ecosystem. Many Graphistry deployments work fine with only Graphistry on one GPU and using its connectors to CPU infrastructure, and horizontally or vertically scaling as the number of analysts increases. However, as RAPIDS matures and is able to do more, the new usage modes benefit from a different kind of planning.

Graphistry and RAPIDS are increasingly able to connect to any data source, including flatfiles, and support multi-GPU + multi-node compute over them. Instead of planning in terms of the number of concurrent analyst sessions as in standard Graphistry, it becomes useful to consider the total raw data size or the size of the typical data extract. GPUs work best when all the data fits in GPU memory, or for larger datasets, when big network connections or multiple smaller ones are made available. It is still the early days, so we recommend discussing with a member of our team.