Skip to content

sizing.md

Latest commit

 

History

History
63 lines (46 loc) · 2.52 KB

sizing.md

File metadata and controls

63 lines (46 loc) · 2.52 KB
Raw

Sizing — profiles mapped to instance types

DarkMoon is a single-node appliance: you scale up (a bigger instance), not out. Pick the profile that matches the workload you actually run — assets × agents × tool executions × parallelism × campaign complexity × evidence retention — and treat the figures as sizing estimates with margin, not hard floors.

Profiles → instance size

Profile vCPU / RAM AWS GCP Azure OVH
Minimum 2 / 8 t3.large e2-standard-2 Standard_B2ms b2-15
Standard 4 / 16 t3.xlarge e2-standard-4 Standard_D4s_v5 b2-30
Performance 8 / 32 t3.2xlarge e2-standard-8 Standard_D8s_v5 b2-60
Industrial 16 / 64 m6i.4xlarge e2-standard-16 Standard_D16s_v5 b2-120

Which to pick

  • Minimum — a single low-parallelism campaign against a handful of hosts.
  • Standard — one campaign at normal parallelism. A good default.
  • Performance — high concurrency or larger campaigns.
  • Industrial — sustained edge node with long evidence retention.

Architecture: amd64 only

Use amd64 (x86-64) instance types. arm64 is experimental — do not select arm64 instances (Graviton, Ampere/Axion, Dpsv5/b3 ARM, etc.) for a production node. Every instance type in the table above is amd64.

Disk

Provision an SSD/NVMe root or data volume sized to your evidence retention:

  • Minimum: ~40 GB SSD
  • Standard: ~80 GB SSD
  • Performance: ~160 GB NVMe
  • Industrial: ~250 GB+ NVMe

Disk is where campaign evidence accumulates. darkmoon doctor warns at ≥90% used.

Local AI adds memory

A GPU is only for Local inference (on-node model). Connected and Private modes do inference off the node and never need a GPU. If you run Local mode, add memory (RAM, or VRAM if you use a GPU) on top of the base profile, per model size:

Model size Extra memory
7B +8 GB
13B +16 GB
33B +32 GB

So a Standard node (4 / 16) running a 13B local model wants roughly 16 + 16 = 32 GB — step up to a Performance-class instance, or add a GPU with enough VRAM. See AI modes.

Notes

  • These figures are engineering estimates derived from component specs and existing data, not a fresh execution benchmark.
  • Start one profile above your best guess if you are unsure; it is cheaper than re-provisioning (and re-provisioning consumes a license slot — see Doctor & lifecycle).