SIM-PIPE is a state of the art data pipeline system that currently enables step-level energy and performance monitoring of containerized big data and AI pipelines. However, evaluating how different hardware platforms (e.g., CPUs, GPUs, FPGAs, TPUs, edge devices) affect pipeline sustainability often requires physical access to diverse hardware. By integrating hardware emulation into SIM-PIPE, researchers and practitioners can experiment with “what-if” hardware scenarios without requiring large hardware testbeds. The thesis will extend SIM-PIPE with a hardware emulation layer that allows users to evaluate pipelines on virtualized or emulated hardware targets. This includes integrating existing emulation frameworks (e.g., QEMU for CPU architectures, GPGPU-Sim for GPUs, or edge hardware emulators) into SIM-PIPE’s monitoring loop. The student will design mechanisms to redirect pipeline execution traces to emulators, collect performance/energy data, and compare emulated results against real measurements for calibration. Work to be done:
- Literature & Tool Survey
- Study hardware emulation frameworks (QEMU, gem5, GPGPU-Sim, FPGA emulators).
- Review simulation vs. real hardware validation studies in HPC/AI.
- Requirement Analysis
- Identify which SIM-PIPE measurement components (e.g., Carbontracker, Scaphandre) can be adapted for emulated execution.
- Define use cases (e.g., “run pipeline step on emulated ARM CPU vs. x86 CPU”).
- Design & Integration
- Extend SIM-PIPE’s execution model to support redirecting steps to an emulated hardware backend.
- Implement interfaces for performance/energy data retrieval from emulators.
- Prototype Implementation
- Integrate at least one CPU emulator (QEMU or gem5) into SIM-PIPE.
- Optionally add GPU emulation with GPGPU-Sim for comparison.
- Evaluation
- Run representative pipelines (ETL, ML training, log analysis) on both physical and emulated hardware.
- Compare energy/performance metrics: accuracy of emulation, overhead, and scalability.
- Analysis
- Discuss trade-offs: fidelity vs. execution time, portability vs. precision.
- Propose calibration strategies for aligning emulator outputs with real measurements.