That '8GB' is not 8GB: virtio balloon shows how cloud memory oversubscription works
joshalbrecht · x · 2026-09-25
The author explains the mechanics behind cloud memory oversubscription: providers run a virtio balloon, so when you try to use more memory you might get it—slowly—which is not at all the same as actually having 8GB of RAM. A sharp technical point on the gap between advertised memory configs and real availability under load.
More from Infra
- LunarG's KosmicKrisp passes Vulkan 1.4 CTS on Apple Silicon, 3x faster since 1.3 — ssh4net · 2026-09-25
- ServingStudio launches as a simulation workbench for optimizing LLM serving systems — CShorten30 · 2026-09-25
- faster-qwen3-tts brings quantized Qwen3-TTS to Apple Silicon via GGML — andimarafioti · 2026-09-25
- Why KV Cache Eats GPU Memory: The Hidden Cost of Context Length and Concurrency — techNmak · 2026-09-25
- Running local agents on vLLM: only ~8 agents with 32k context each — what use cases actually help? — Ambitious_Fold_2874 · 2026-09-25
- Jensen Huang Pushes Back on AI Doom: "AI Is Software, It's Math" — DavidLinthicum · 2026-09-25