How uw-mad-dash/eva works
Main Components
• Global Storage: Represent the “Global Storage” element (Amazon S3) which holds the job artifacts. This component is accessed by both the master and workers.
• Master Node: This is the central controller that manages job submissions, scheduling, and cloud provisioning. Include submodules such as the scheduler, command, and job management (found in the master/ directory and associated RPC modules).
• Worker Nodes: These execute the jobs. They interact with global storage (using the storage_manager) and follow commands from the master.
• Cloud Provisioner: This component (under src/cloud_provisioner) interacts with AWS EC2 APIs to launch and shut down worker nodes based on resource demand.
• Experiment and Simulation Drivers: Represent the simulation components (e.g. simulation, simulator, experiment_driver_200.py, experiment_driver_full.py) that are used for evaluation of scheduling on both partial and full traces from Alibaba.
• Job Artifacts & Workloads: Represent the directory of job workloads (e.g. workloads/a3c, workloads/resnet18, etc.) which house the Dockerfiles and job-specific configurations used to containerize and execute the jobs.
• EVAIterator: This component (in the eva_iterator directory) that is embedded in user jobs in order to track throughput and interference; it communicates with the master/worker processes via RPC.
Relationships and Interactions
• For simulation experiments, illustrate that a separate driver component (simulation/ and simulator/) interacts with the scheduling logic to run trace-based experiments, generating cost metrics and performance reports.
Architectural Patterns and Design Principles
• The system follows a modular and layered design:
– Separation of Concerns: Cloud provisioning, scheduling, and actual job execution are clearly separated into different modules.
– RPC-based Communications: Many components communicate via defined RPC interfaces (see src/rpc and associated protobuf files).
– Containerized Workloads: Many workload directories include Dockerfiles, which indicates container-based job execution.
• Use of configuration files (e.g., eva_config.json, ec2_config.json, s3_config.json) that govern system behavior; these should be shown as configuration “settings” feeding into the Master Node and Cloud Provisioner.
Relevant Technologies and Frameworks
• AWS-specific components (EC2 for compute, S3 for storage).
• Containerization technology (Docker) used for workload execution.
• Reference Python-based execution and scheduling components.
• RPC mechanisms (often generated from protobuf definitions) that enable communication across modules.
Must-Have Elements
• Each component should have a clear label (for example “Master Node – Scheduling & Job Management”, “Worker Node – Execution”, “Cloud Provisioner – AWS EC2 APIs”, “Global Storage – Amazon S3”).
– Command flow from Master to Workers.
– Data flow from Global Storage to Worker Nodes and vice versa.
– Control flow from Cloud Provisioner to launching new Worker Nodes.
• Optionally, include configuration files as a supporting block that influences the operation of both the master (via eva_config.json, ec2_config.json) and provisioning components.
1. Job submission (via eva_submit.py or job_submission_driver.py)
2. Master scheduling and decision making
3. Worker node launching (via Cloud Provisioner)
4. Job artifact retrieval and container execution on the Worker
5. Synchronization back to Global Storage