How to Kickstart Your CS Research Journey and Find LLM Serving Ideas

At the end of the year the author reflects on a half‑year research journey, aiming to record insights and exchange experiences with peers. The motivation is twofold: to think ahead about learning and to document progress inspired by a popular Zhihu year‑end summary.

How to Start Research and Find Ideas?

The author began by reading a series of posts curated by Dr. Xia Zhao, which provide an overview of CS system research and highlight influential papers from conferences such as OSDI, ASPLOS, ATC, and SOSP. Subscribing to the Arxiv daily digest (covering AI, hardware architecture, distributed computing, operating systems, and software engineering) helped stay updated on cutting‑edge work, especially in HPC and serverless computing.

After gaining a broad sense of the field, the author focused on reading many papers, taking notes, and summarizing key points. Particular interest was drawn to LLM systems (LLM SYS) after the release of the Llama 3 technical report, leading to deeper exploration of papers on CPU‑GPU joint inference and other serving optimizations.

Choosing a Specific Direction

Seeing that traditional cloud scheduling research was becoming saturated, the author selected LLM serving as a promising intersection of resource scheduling and large‑language‑model deployment. Attending Stanford ML seminars and UC‑Berkeley LLM‑agents talks further refined the research focus.

Designing a system revealed the complexity of integrating multi‑disciplinary knowledge. The author consulted experts across domains, built a prototype, and presented it to the advisor, who suggested splitting the work into multiple papers rather than a single large one. This highlighted the importance of simplifying the problem, extracting core contributions, and iterating between reading papers and building prototypes.

Collaboration Is Essential

The narrative stresses that research is not a solitary endeavor. Effective communication with advisors, lab mates, and external collaborators prevents burnout and misdirection. The author acknowledges that without a supportive network, exploring a novel LLM‑serving direction would be far more challenging.

Conclusion

The author hopes the reflections aid newcomers to academic research, inviting comments and further discussion while acknowledging that the writing itself still needs improvement.