Compound AI Systems, integrating multiple interacting components like models, retrievers, and external tools, have emerged as essential for addressing complex AI tasks. However, current implementations suffer from inefficient resource utilization due to tight coupling between application logic and execution details, a disconnect between orchestration and resource management layers, and the perceived exclusiveness between efficiency and quality. We propose a vision for resource-efficient Compound AI Systems through a declarative workflow programming model and an adaptive runtime system for dynamic scheduling and resource-aware decision-making. Decoupling application logic from low-level details exposes levers for the runtime to flexibly configure the execution environment and resources, without compromising on quality. Enabling collaboration between the workflow orchestration and cluster manager enables higher efficiency through better scheduling and resource management. We are building a prototype system, called Murakkab, to realize this vision. Our preliminary evaluation demonstrates speedups up to $\sim 3.4\times$ in workflow completion times while delivering $\sim 4.5\times$ higher energy efficiency, showing promise in optimizing resources and advancing AI system design.

Performance testing aims at uncovering efficiency issues of software systems. In order to be both effective and practical, the design of a performance test must achieve a reasonable trade-off between result quality and testing time. This becomes particularly challenging in Java context, where the software undergoes a warm-up phase of execution, due to just-in-time compilation. During this phase, performance measurements are subject to severe fluctuations, which may adversely affect quality of performance test results. However, these approaches often provide suboptimal estimates of the warm-up phase, resulting in either insufficient or excessive warm-up iterations, which may degrade result quality or increase testing time. There is still a lack of consensus on how to properly address this problem. Here, we propose and study an AI-based framework to dynamically halt warm-up iterations at runtime. Specifically, our framework leverages recent advances in AI for Time Series Classification (TSC) to predict the end of the warm-up phase during test execution. We conduct experiments by training three different TSC models on half a million of measurement segments obtained from JMH microbenchmark executions. We find that our framework significantly improves the accuracy of the warm-up estimates provided by state-of-practice and state-of-the-art methods. This higher estimation accuracy results in a net improvement in either result quality or testing time for up to +35.3% of the microbenchmarks. Our study highlights that integrating AI to dynamically estimate the end of the warm-up phase can enhance the cost-effectiveness of Java performance testing.

Diffusion models, such as Stable Diffusion (SD), offer the ability to generate high-resolution images with diverse features, but they come at a significant computational and memory cost. In classifier-free guided diffusion models, prolonged inference times are attributed to the necessity of computing two separate diffusion models at each denoising step. Recent work has shown promise in improving inference time through distillation techniques, teaching the model to perform similar denoising steps with reduced computations. However, the application of distillation introduces additional memory overhead to these already resource-intensive diffusion models, making it less practical. To address these challenges, our research explores a novel approach that combines Low-Rank Adaptation (LoRA) with model distillation to efficiently compress diffusion models. This approach not only reduces inference time but also mitigates memory overhead, and notably decreases memory consumption even before applying distillation. The results are remarkable, featuring a significant reduction in inference time due to the distillation process and a substantial 50% reduction in memory consumption. Our examination of the generated images underscores that the incorporation of LoRA-enhanced distillation maintains image quality and alignment with the provided prompts. In summary, while conventional distillation tends to increase memory consumption, LoRA-enhanced distillation offers optimization without any trade-offs or compromises in quality.

Community detection is a well-studied problem with applications in domains ranging from networking to bioinformatics. Due to the rapid growth in the volume of real-world data, there is growing interest in accelerating contemporary community detection algorithms. However, the more accurate and statistically robust methods tend to be hard to parallelize. One such method is stochastic block partitioning (SBP) - a community detection algorithm that works well on graphs with complex and heterogeneous community structure. In this paper, we present a sampling-based SBP (SamBaS) for accelerating SBP on sparse graphs. We characterize how various graph parameters affect the speedup and result quality of community detection with SamBaS and quantify the trade-offs therein. To evaluate SamBas on real-world web graphs without known ground-truth communities, we introduce partition quality score (PQS), an evaluation metric that outperforms modularity in terms of correlation with F1 score. Overall, SamBaS achieves speedups of up to 10X while maintaining result quality (and even improving result quality by over 150% on certain graphs, relative to F1 score).

Statistical machine learning often uses probabilistic algorithms, such as Markov Chain Monte Carlo (MCMC), to solve a wide range of problems. Many accelerators are proposed using specialized hardware to address sampling inefficiency, the critical performance bottleneck of probabilistic algorithms. These accelerators usually improve the hardware efficiency by using some approximation techniques, such as reducing bit representation, truncating small values to zero, or simplifying the Random Number Generator (RNG). Understanding the influence of these approximations on result quality is crucial to meeting the quality requirements of real applications. Although a common approach is to compare the end-point result quality using community-standard benchmarks and metrics, we claim a probabilistic architecture should provide some measure (or guarantee) of statistical robustness. This work takes a first step towards quantifying the statistical robustness of specialized hardware MCMC accelerators by proposing three pillars of statistical robustness: sampling quality, convergence diagnostic, and goodness of fit. Each pillar has at least one quantitative metric without the need to know the ground truth data. We apply this method to analyze the statistical robustness of an MCMC accelerator proposed by previous work, with some modifications, as a case study. The method also applies to other probabilistic accelerators and can be used in design space exploration.

Web search engines respond to a query by returning more results than can be reasonably reviewed. These results typically include the title, link, and snippet of content from the target link. Each result has the potential to be useful or useless and thus reviewing it has a cost and potential benefit. This paper studies the behavior of a rational agent in this setting, whose objective is to maximize the probability of finding a satisfying result while minimizing cost. We propose two similar agents with different capabilities: one that only compares result snippets relatively and one that predicts from the result snippet whether the result will be satisfying. We prove that the optimal strategy for both agents is a stopping rule: the agent reviews a fixed number of results until the marginal cost is greater than the marginal expected benefit, maximizing the overall expected utility. Finally, we discuss the relationship between rational agents and search users and how our findings help us understand reviewing behaviors.