pipascope 0.0.1

No project description provided

PipaScope: A Dataset for CPU Microarchitecture Performance Characterization

PipaScope – Observe the pulse of performance, one cycle at a time.
An open dataset initiative for microarchitectural behavior analysis, led by ZJU-SPAIL.

🍈 About the Name: PIPA & PipaScope

PIPA (Progressive Intelligent Performance Analytics) draws inspiration from loquat (枇杷), a fruit native to Zhejiang, China. Its lifecycle—tree (collecting), flower (analysis), and fruit (conclusion)—mirrors the performance engineering pipeline.

PipaScope extends this metaphor as the observational lens into the microarchitectural world. Just as the loquat tree absorbs nutrients from the soil, PipaScope captures low-level performance telemetry from real workloads, enabling deep insight into CPU behavior.

This dataset serves as the foundational "soil" for training automated performance diagnosis systems.

🏫 Project Ownership

PipaScope is currently led and maintained by the System Performance Analytics and Intelligence Lab (ZJU-SPAIL) at Zhejiang University.
It is part of ongoing research into systematic performance characterization and bottleneck analysis.

🎯 Focus: Microarchitectural Behavior

PipaScope is designed to support research in CPU microarchitecture performance characterization, with a focus on:

• Instruction per Cycle (IPC) degradation
• Cache miss patterns (L1/L2/LLC)
• Memory bandwidth saturation
• Frontend/backend stalls
• Branch misprediction penalties
• TLB pressure

The goal is to build a high-quality, version-controlled dataset that enables reproducible analysis and lays the foundation for automated bottleneck identification.

🧩 Data Sources

The dataset includes performance profiles from:

• SPEC CPU 2017 (both integer and floating-point benchmarks)
• Real-world applications, starting with RocksDB

Each workload is executed under diverse configurations (input sets, system settings, compiler flags) and on multiple hardware platforms (Intel/Arm) to capture a wide range of microarchitectural behaviors.

🛠️ Data Collection

All data is collected using standardized tools and methodologies:

• perf (Linux Performance Events) for hardware counter sampling
• Custom run scripts for SPEC CPU 2017 and real-world applications
• Metric derivation based on PIPA-SHU principles (multiplexing-aware counter aggregation)

All data is versioned using Git LFS to support large file storage and traceability.

📌 Status

This project is in the early development phase.
The dataset is actively being built by ZJU-SPAIL members.
Public access is read-only; contributions are not currently accepted.

Documentation and tooling will be expanded as the dataset matures.

---

“PipaScope: where data grows like fruit, and insight blossoms from observation.”