Skip to content

Releases: AutoCookies/gslice

gslice V1.0.0

20 Feb 04:19
@AutoCookies AutoCookies
gslice-v1.0.0
801495f

Choose a tag to compare

View all tags
gslice V1.0.0 Latest
Latest

🚀 GPU Slice v1.0.0

Production-Ready Memory Quota Enforcement for CUDA Workloads (Software-Based, No MIG Required)

🎯 What This Release Delivers

GPU Slice v1.0.0 introduces a software-enforced VRAM isolation layer for CUDA workloads running on Linux — without requiring NVIDIA MIG or expensive datacenter GPUs.

This release provides:

  • ✅ Deterministic per-session VRAM quota enforcement

  • ✅ LD_PRELOAD-based CUDA interception

  • ✅ Crash-safe quota recovery

  • ✅ Session TTL & expiration handling

  • ✅ Local IPC authentication

  • ✅ Prometheus metrics

  • ✅ Structured audit logging

  • ✅ Production-ready CLI workflow

  • ✅ Deterministic benchmark suite

  • ✅ Clean Architecture (Hexagonal) design

This is not a prototype.
This is a hardened, tested v1.0.0 release.

🧠 Problem Statement

Most consumer GPUs (RTX 3090 / 4090) lack hardware partitioning (MIG).
When multiple AI workloads share a GPU:

  • One process can exhaust VRAM

  • OOM errors cascade unpredictably

  • No isolation between tenants

  • No enforcement boundaries

GPU Slice provides:

Software-level VRAM isolation without modifying the NVIDIA driver.

🏗 Architecture Overview

High-Level Flow

flowchart LR
    UserProcess -->|LD_PRELOAD| Interceptor
    Interceptor -->|IPC (UDS)| ControlPlane
    ControlPlane --> Store
    ControlPlane --> Metrics

Allocation Lifecycle

sequenceDiagram
    participant App
    participant Interceptor
    participant ControlPlane

App->>Interceptor: cudaMalloc(size)
Interceptor->>ControlPlane: reserve(session, size)
ControlPlane-->>Interceptor: allow / deny
Interceptor->>CUDA: call real allocation

Crash Recovery Loop

flowchart TD

Tick --> ScanSessions

ScanSessions --> CheckPID

CheckPID -->|Dead| ReclaimBytes

🔐 Security Model

  • Unix Domain Socket (local-only IPC)

  • Shared secret token authentication

  • Constant-time token comparison

  • Fail-closed behavior on allocation if IPC unavailable

  • TTL-based session expiration

  • PID-based orphan allocation recovery

🛠 What’s Included

Control Plane (Go)

  • Session management

  • Quota accounting

  • Allocation registry

  • Crash recovery loop

  • TTL expiration

  • Prometheus metrics

  • Audit log (JSON lines)

Interceptor (C, LD_PRELOAD)

  • Hooks:

    • cudaMalloc

    • cudaFree

    • cudaMallocManaged

    • cudaMallocPitch

  • Thread-safe allocation tracking

  • IPC enforcement

  • Fail-closed allocation policy

  • No external C dependencies

CLI

gpuslice run --limit 128MB -- python app.py

Automatically:

  • Allocates session

  • Injects LD_PRELOAD

  • Sets env vars

  • Handles signals

  • Releases session on exit

📊 Benchmark Results (v1.0.0)

(Example structure — actual values generated via bench suite)

Allocation Overhead

Mode Avg ns/op Overhead
Baseline 120ns —
With Slicer 180ns +50%

Stress Test

  • 10 concurrent processes

  • 100 allocations each

  • Quota enforcement correct

  • No leak after crash

  • Recovery within 2s

Overhead is predictable and bounded.

🧪 Reliability Features

Crash Recovery

If a process dies without freeing memory:

  • PID detected via /proc

  • Allocations reclaimed automatically

  • No permanent quota leak

Server Restart Safety

Allocation registry persisted.
Recovery replays on restart.

Fail-Closed Enforcement

If IPC is unreachable:

  • Allocation denied

  • Prevents runaway memory usage

📈 Observability

Metrics exposed at /metrics:

  • gpuslice_sessions_active

  • gpuslice_used_bytes_total

  • gpuslice_alloc_events_total

  • gpuslice_denied_alloc_total

  • gpuslice_recovered_bytes_total

Structured logs include:

  • session_id

  • pid

  • operation

  • bytes

  • result

  • error_code

Optional audit log file supported.

🚀 Installation

make build
make demo
make bench

Environment variables:

GPUSLICE_SESSION
GPUSLICE_IPC_SOCK
GPUSLICE_IPC_TOKEN
GPUSLICE_DEBUG

🧩 Production Usage Example

export GPUSLICE_IPC_TOKEN=supersecret

./gpusliced &


gpuslice run --limit 256MB -- python train.py

⚠️ Limitations (Intentional)

  • Memory quota only

  • No compute scheduling

  • No hardware partitioning

  • No multi-node coordination

  • Linux only

This is a memory isolation layer — not a GPU hypervisor.

🗺 Roadmap (Post v1.0.0)

Future (separate milestones):

  • Compute fairness (research required)

  • Kubernetes device plugin

  • Multi-node quota federation

  • Optional billing integration

No premature scope expansion.

🧱 Design Principles

  • Clean Architecture (Hexagonal)

  • Domain purity

  • Deterministic enforcement

  • Fail-safe behavior

  • Minimal C surface

  • No hidden global state

  • No external C dependencies

  • Reproducible builds

🏁 Release Summary

GPU Slice v1.0.0 is:

  • A stable

  • Tested

  • Deterministic

  • Production-hardened

  • Memory-only GPU isolation layer

Built for real workloads, not demos.

    Assets 2
    Loading