DyPamear is an efficient Dynamic Graph Pattern Mining (DGPM) system for the UPMEM Processing-in-Memory (PIM) architecture.

The goal of DyPamear is to accelerate incremental pattern-aware execution of DGPM using UPMEM DPUs. It integrates asynchronous pipelining, bitmap-based set intersection, and other optimizations tailored for near-data processing architecture.

• Load-aware task distribution for balanced execution across thousands of DPUs.
• Asynchronous loader–worker pipeline using WRAM FIFO.
• Bitmap-based set intersection acceleration on DPUs.
• Lightweight performance profiling and cycle analysis tools.

DyPamear-AE/
├── host/         # Host-side logic (C)
├── dpu/          # DPU-side programs (C for UPMEM)
├── python_tool/  # Python scripts for preprocessing and profiling
├── include/      # Shared headers
├── makefile      # Compilation rules
└── README.md     # Project description

• Linux environment
• UPMEM SDK v2025.1.0.
• GNU Make, C compiler (e.g., gcc)
• Python ≥ 3.8 (for analysis scripts)

To match a pattern within a graph, run:

make clean
GRAPH=<graph_name> UPDATES=<updated_edges> PATTERN=<pattern_name> make test

Example:

GRAPH=WK UPDATES=WK_updates PATTERN=CLIQUE3 make test

💡 The available values for GRAPH and PATTERN are defined in include/common.h.
To add new graphs or patterns, modify common.h and recompile.

DyPamear supports flexible definitions of graph inputs and matching patterns.

All configuration entries are defined in include/common.h.

1. Place your input graph (in CSR binary format) into the ./data/ directory.
2. Add a macro definition in include/common.h:

#if defined(AM0312)
#define DATA_NAME "amazon0312_adj"
#define N (1<<20)
#define M (1<<23)
#endif

3. Build and test:

GRAPH=WK UPDATES=WK_updates PATTERN=CLIQUE3 make test

1. Define a new macro for your pattern kernel in include/common.h

#elif defined(TELE5)
#define KERNEL_FUNC tele5
#define PATTERN_NAME "tele5"
#endif

2. Implement the kernel function in dpu/ directory (e.g., in TELE5.c or new source file).
3. Build and run:

GRAPH=WK UPDATES=WK_updates PATTERN=TELE5 make test

DyPamear is designed to scale from hundreds to tens of thousands of DPUs.

To run DyPamear on a specific number of DPUs:

GRAPH=WK UPDATES=WK_updates PATTERN=CLIQUE3 EXTRA_FLAGS="-DV_NR_DPUS=5120" make test

To automatically benchmark DyPamear from 640 to 20,480 DPUs:

GRAPH=WK UPDATES=WK_updates PATTERN=CLIQUE3 make test_sc

This script:

• Compiles DyPamear with various DPU counts.
• Runs the benchmark for each configuration.

Analyzes CSR binary and outputs graph statistics:

python3 python_tool/analyze_csr_graph.py input/graph.bin

Outputs include:

• Number of nodes and edges
• Degree distribution (min/avg/max)

Visualizes DPU-level workload distribution:

python3 python_tool/show_cycle.py result.txt

• Left plot: Max cycle per DPU
• Right plot: Task count per DPU