GPU Graph Memory Optimization Lab

A CUDA-based benchmark framework that studies how GPU memory allocation strategies impact graph analytics workloads.

The project implements PageRank on real-world graph datasets and compares:

1. cudaMalloc
2. cudaMallocAsync
3. Custom GPU Memory Pool
4. Unified Memory

A CUDA-based benchmark framework that studies GPU memory allocation strategies for large-scale graph analytics workloads using PageRank and graph traversal algorithms.

Real-world motivation Companies process huge graphs:

Examples:

Social networks → user connections Fraud detection → transaction networks Recommendation systems → user-item graphs Scientific computing → mesh graphs Cybersecurity → network graphs A graph workload has a unique GPU challenge:

memory access dominates computation.

Unlike matrix multiplication:

A × B where memory access is predictable:

Graphs look like:

Node 0 | +---- Node 45 | +---- Node 9203 | +---- Node 17 The GPU constantly jumps around memory.

Core idea All four versions run:

SAME GRAPH

SAME PAGE RANK ALGORITHM

SAME GPU KERNEL

ONLY MEMORY ALLOCATION CHANGES This is important.

Your experiment is controlled.

This project investigates:

How does GPU memory allocation strategy affect graph analytics performance?

---

Workload

The benchmark uses PageRank.

PageRank repeatedly updates node importance scores by distributing rank values across graph edges.

What the finished project does You will run:

./build/graph_benchmark --allocator cudaMalloc then:

./build/graph_benchmark --allocator cudaMallocAsync then:

./build/graph_benchmark --allocator memory_pool then:

./build/graph_benchmark --allocator unified_memory All four will run:

same graph same PageRank kernel same number of iterations same GPU

Only memory management changes.

Benchmark Metrics

The framework measures:

Runtime total execution time average iteration time Memory initial GPU memory peak memory usage Allocation allocation count allocated bytes allocation overhead GPU Metrics