About Me

📚 Education #

Boston University ↗
📝 PhD Student in Computer Science
🗓 Sep 2024 – Present · 📍 Boston, MA

Boston University ↗
🎓 Master of Science in Computer Science — GPA: 3.96/4.0
🗓 Sep 2022 – May 2024 · 📍 Boston, MA

South China University of Technology ↗
🎓 Bachelor of Engineering in Software Engineering
🗓 Sep 2017 – Jun 2021 · 📍 Guangzhou, China

🧰 Work & Internship #

Intel Asia-Pacific Research & Development Ltd — AI Framework Engineer ↗
🗓 Aug 2021 – Aug 2022 · 📍 Shanghai, China

• Contributed to the oneDNN Graph Compiler with fusion, loop rescheduling, inlining, and SIMD intrinsics, delivering measurable speedups in deep learning workloads across AVX512/AVX512-BF16 platforms.
• Implemented a range of elementwise and graph operators including normalization and VNNI reordering, developed graph patterns such as MHA to identify and analyze performance bottlenecks.

Tencent – WeChat Group — Software Engineer Intern ↗
🗓 Jun 2020 – Sep 2020 · 📍 Guangzhou, China

• Built an internal test case management system using Python Tornado + React.
• Developed Grafana dashboards for internal statistical data visualization.

📖 Research & Publications #

[Paper] RingSampler: GNN Sampling on Large-Scale Graphs with io_uring — HotStorage ’25 ↗
🗓 Jul 2025 (Second Author) · 🔗 ACM Link

• Introduced an io_uring-based out-of-core sampling method for graphs larger than memory.
• Fully utilized CPU resources while achieving near in-memory sampling performance.

[Poster] GNN Training with Graph Summarization — MLSys ’25 ↗
🗓 May 2025

• Explored graph summarization as preprocessing to shrink graph size before training.
• Demonstrated lower sampling time vs. random-access baseline and significant PCIe bandwidth savings.

SmartSSD Code Repository ↗
🗓 Dec 2024 ·  GitHub

• Published an evolved implementation of SmartSSD-based GNN sampling.
• Refactored host-side design for clearer results and usability.

[Paper] In Situ Neighborhood Sampling for Large-Scale GNN Training — DaMoN ’24
🗓 Jun 2024 · 🔗 ACM Link

• Proposed a SmartSSD-based streaming sampler for efficient GNN training.
• Demonstrated lower sampling time vs. random-access baseline and significant PCIe bandwidth savings.

[Master’s Thesis] Accelerating Large-Scale GNN Training with Programmable SSDs
🗓 May 2024 · 📍 Boston University · 🔗 Thesis Link

• Designed FPGA kernels with Vitis HLS to offload computation onto SSD and reduce PCIe transfers.
• Compared preprocessing costs, file layouts, and sampling speeds on Papers100M and Yahoo datasets.

🔬 Projects #

Advanced Matrix Multiplication — Strassen Algorithm ↗
🗓 Feb 2023 – May 2023 · 📍 Boston, MA

• Implemented and optimized the Strassen algorithm for fast matrix multiplication.
• Leveraged OpenMP parallelism and SIMD intrinsics to accelerate computation.
• Conducted detailed performance analysis and compared against baseline algorithms.

Fine-Tuning Pretrained Vision Transformers ↗
🗓 Feb 2023 – May 2023 · 📍 Boston, MA

• Fine-tuned NAT, DiNAT, MaxViT, and DaViT models on the iNaturalist dataset.
• Achieved competitive accuracy compared to state-of-the-art models.
• Analyzed architectural differences and compared fine-tuning vs. training from scratch.

Online Signature Verification (CNN + GRU) ↗
🗓 Jan 2021 – June 2021 · 📍 Guangzhou, China

• Extracted static features from signature images via a CNN-based autoencoder.
• Captured dynamic features from signature trajectories using a GRU-based autoencoder.
• Applied early/late fusion and a CNN-based Siamese network to verify authenticity.

Parallel Group-By Operation (MPI + OpenMP) ↗
🗓 Mar 2020 - June 2020 · 📍 Guangzhou, China

• Simulated the “Group-By” operation of databases on large datasets.
• Used OpenMP for node-level parallelism and MPI for distributed cross-node processing.
• Achieved multi-fold efficiency improvements over serial implementations.

Short Video Copyright Detection ↗
🗓 Sep 2019 – Nov 2019 · 📍 Guangzhou, China

• Designed an approximate detection system to find short clips within long videos.
• Extracted features with SIFT and indexed them using FAISS vector database.
• Applied hierarchical matching + sliding window for accurate detection within 5s tolerance.

UWP-Based Property Management System ↗
🗓 Jun 2019 – Sep 2019 · 📍 Guangzhou, China

• Built a frontend with C++ and UWP and a backend with Python/Django.
• Designed and implemented RESTful APIs for client-server communication.
• Delivered a complete property management solution for internal use.

🏆 Honors & Awards #

• 1st Scholarship, SCUT — Awarded for academic excellence (Oct 2020) ↗
• 2nd Prize, CCF NOIP (Nov 2015) — Prestigious algorithm competition akin to ACM-ICPC ↗

📱 Contact #

• 📧 Email: i [at] syh.one
• 💻 GitHub: github.com/Souukou
• 🔗 LinkedIn: linkedin.com/in/yuhangsong