The ever increasing scale of modern data center demands more effective optimizations, as even a small percentage of performance improvement can result in a significant reduction in data-center cost and its environmental footprint. However, the diverse set of workloads running in data centers also challenges the scalability of optimization solutions. Profile-guided optimization (PGO) is a promising technique to improve application performance. Sampling-based PGO is widely used in data-center applications due to its low operational overhead, but the performance gains are not as substantial as the instrumentation-based counterpart. The high operational overhead of instrumentation-based PGO, on the other hand, hinders its large-scale adoption, despite its superior performance gains. In this paper, we propose CSSPGO, a context-sensitive sampling-based PGO framework with pseudo-instrumentation. CSSPGO offers a more balanced solution to push sampling-based PGO performance closer to instrumentation-based PGO while maintaining minimal operational overhead. It leverages pseudo-instrumentation to improve profile quality without incurring the overhead of traditional instrumentation. It also enriches profile with context-sensitivity to aid more effective optimizations through a novel profiling methodology using synchronized LBR and stack sampling. CSSPGO is now used to optimize over 75% of Meta's data center CPU cycles. Our evaluation with production workloads demonstrates 1%-5% performance improvement on top of state-of-the-art sampling-based PGO.

Profile-guided optimization (PGO) is an important component in modern compilers. By allowing the compiler to leverage the program’s dynamic behavior, it can often generate substantially faster binaries. Sampling-based profiling is the state-of-the-art technique for collecting execution profiles in data-center environments. However, the lowered profile accuracy caused by sampling fully optimized binary often hurts the benefits of PGO; thus, an important problem is to overcome the inaccuracy in a profile after it is collected. In this paper we tackle the problem, which is also known as profile inference and profile rectification . We investigate the classical approach for profile inference, based on computing minimum-cost maximum flows in a control-flow graph, and develop an extended model capturing the desired properties of real-world profiles. Next we provide a solid theoretical foundation of the corresponding optimization problem by studying its algorithmic aspects. We then describe a new efficient algorithm for the problem along with its implementation in an open-source compiler. An extensive evaluation of the algorithm and existing profile inference techniques on a variety of applications, including Facebook production workloads and SPEC CPU benchmarks, indicates that the new method outperforms its competitors by significantly improving the accuracy of profile data and the performance of generated binaries.