New system manages tasks across multiple Macs
Metal Sandbox uses command scheduling and isolated environments to improve efficiency on Apple Silicon.
Metal Sandbox is a new infrastructure designed for Apple Silicon. It combines command-level scheduling with restricted file access. This system works across interconnected Macs.
Running models on devices requires kernels that use local accelerators well. Metal provides a native GPU API for M-series and A-series chips. These processors use unified memory to support CPU and GPU workflows.
Current sessions alternate between model requests, development commands, and GPU execution. One metal_nbody session showed only 44.0 seconds of GPU activity over 31.6 minutes. This means GPU capacity remains idle for 2.32% of the session.
Metal Sandbox provides each agent with a private development HOME and workspace. It uses Apple's App Sandbox to restrict access to files. This prevents agents from writing to the host user's files or other instances.
The system uses a serial queue for commands that use the GPU. Parallel commands like file editing can run without waiting. If a Mac is too busy, waiting commands can migrate to other Macs.
Experiments on four Mac mini M4 nodes showed gains in command throughput. Under concentrated demand, throughput increased by --. Using sixteen agent traces, the estimated sustained-throughput gain was for four sessions per node.
Dynamic scheduling reduced mean tool waiting by 18.8%. It also reduced batch elapsed time by 22.3%. These results assume phase durations remain unchanged.
An instance is a workspace and environment for an agent session. A node is a Mac that hosts a session or runs commands. Migration is the process of moving a command to a different Mac.
Key facts
- Metal Sandbox achieves -- gains in command throughput on four Mac mini M4 nodes.
- The system reduces mean tool waiting by 18.8% through dynamic scheduling.
- Commands can migrate from a busy Mac to an available Mac in a scheduling pool.