mlsys23-fm4os-2.pdf

Background:

• OS: tasks can be considered sequential decision-making processes where past actions and states of the OS instruct the action at any time

• Existing Methods:

  • DNN to learn policies for specific OS components
  • Statistical and Deep learning methods to generate realistic workloads

  ⇒ remain point solutions that model individual OS components

  ⇒ OS components are not independent of one another

• Goal: model OS components’ relationship to generalize workload to generate well to unseen inputs

• Foundation Model: ML model trained on a large and diverse dataset to understand the general structure of the data and then fine-tuned for specific tasks

  → FM4OS (pre-trained using self-supervised methods on large corpus of OS traces)

Data Sources for FM4OS

• Action logs from OS components → Kernel Logs
  • capture kernel debugging data, hardware events, system events (interrupts / process restarts)
• Resource metrics and Hardware counters → hardware drivers
  • record resource states at pre-defined frequency: CPU, memory, and disk bandwidth utilization, NIC queue length, and hardware counters
• Application workloads

interdependence of OS components can be shown by OS traces

→ use this “natural behavior” of OS to build a foundational model

Tasks for FM4OS

1. Policy Agent :
   1. Make low-level decisions by encompassing tradeoffs that are universally applicable to all OS systems
      1. Fine-tune to make decisions for optimal actions
   2. Policy selection
      1. Models right now cannot match the pace at which OS tasks require
      2. FM4OS can in the meantime focus on choosing the most optimal policy (LRU or LFU) depending on the situation
   3. Challenge:

      1. Composability: policies are also inter-dependent

         ⇒ could jointly fine-tune components

      2. Explainability: goal is to make humans understand OS to a certain degree using the model, so that they can fine-tune themselves

2. Generative Model
   1. Generating traces → training FM4OS using auto-regressive tasks can lead to training the model to generate OS traces that can be used along with the training set
   2. Can generate pathological corner cases → difficult to get irl
   3. Challenge:
      1. quantification of the quality of synthetic samples
3. Predictive Model
   1. Goal is to use FM4OS as encoder of state to predict things about system’s response, future utilization
      1. efficient placement, scheduling, performance, and anomaly detection