Jamilan, S. (CSE) – Profile-guided Compiler Optimizations for Data Center Workloads
Modern applications, such as data center workloads, have become increasingly complex. These applications primarily operate on massive datasets, which involve large memory footprints, irregular access patterns, and complex control and data flows. The processor-memory speed gap, combined with these complexities, can lead to unexpected performance inefficiencies in these applications, preventing them from achieving optimal performance. Considering the complexity and size of data center applications, manually identifying and resolving performance issues is often impractical or impossible. Instead, developing new compiler optimization techniques can be a more effective and scalable solution to boost both performance and energy efficiency. In this thesis, we focus on identifying the root causes that limit the performance of data center workloads. We analyze the limitations of current profile-guided compiler optimization techniques for addressing these performance gaps. Finally, we propose two profile-guided optimization techniques, APT-GET and RIFS, which can be integrated into the LLVM optimization pipeline to deliver further improvements. To hide the long latency of memory accesses, we introduce APT-GET, a profile-guided technique that ensures timely prefetches by leveraging dynamic execution-time information to build a novel analytical model that finds the optimal prefetch distance and injection site based on the collected profile. We study APT-GET across 10 real-world applications and demonstrate that it achieves a speedup of up to 1.98× and an average of 1.30×. To enable runtime value-invariant function specialization to reduce redundant operations, we introduce RIFS, a profile-guided compiler technique that specializes functions based on runtime-invariant call-site-specific argument values. RIFS introduces a novel value-profiling LLVM pass to identify runtime invariant arguments and a subsequent LLVM transformation pass to generate specialized function variants tailored to these value profiles. To efficiently select among potentially thousands of specialization candidates, we develop a predictive cost model that estimates each candidate’s performance benefit before code generation. RIFS achieves an average speedup of 5.3% and an instruction reduction of 2.5% over the LLVM -O3+PGO baseline across 12 real-world applications.
Host: Saba Jamilan, Ph.D. Candidate, Computer Science and Engineering
Advisor: Heiner Litz
Zoom- https://ucsc.zoom.us/j/95818759324?pwd=rdaS7G1V7O6faRhNOgFyq1OR50eSLK.1
Passcode- 652917
