Prompted by the rise of artificial intelligence (AI) or machine learning (ML), more serverless workloads demand efficient GPU support. Recent years have witnessed a shift of interest from weak isolation-based methods, such as Multi-Process Service (MPS), to strong isolation-based methods, such as Multi-Instance GPU (MIG), for GPU support on serverless platforms, thanks to concerns about performance interference and security. The current MIG-based solution for serverless computing is, however, subject to severe GPU resource fragmentation and under-utilization. This paper identifies the reason as the gap between current MIG supports in serverless computing and the rigid constraints in MIG (re)configurations. It proposes FluidFaaS, a solution that enables flexible MIG management for serverless computing. Through a novel programming system support, FluidFaaS enables fine-grained resource assignment to the components within a serverless function, based on which, it equips the invokers with runtime support that constructs pipelines on MIGs on the fly for a serverless function. The innovations, along with a hotness-aware eviction-based time sharing of MIG slices, significantly reduce GPU resource fragmentation and enhance system throughput. Evaluations demonstrate that FluidFaaS outperforms the state-of-the-art solutions by 25%-75% in throughput while achieving up to 90% higher SLO hit rates in various workloads.

Input-centric program optimization aims to optimize code by considering the relations between program inputs and program behaviors. Despite its promise, a long-standing barrier for its adoption is the difficulty of automatically identifying critical features of complex inputs. This paper introduces a novel technique, reductive analysis through compiler-guided Large Language Models (LLMs) , to solve the problem through a synergy between compilers and LLMs. It uses a reductive approach to overcome the scalability and other limitations of LLMs in program code analysis. The solution, for the first time, automates the identification of critical input features without heavy instrumentation or profiling, cutting the time needed for input identification by 44× (or 450× for local LLMs), reduced from 9.6 hours to 13 minutes (with remote LLMs) or 77 seconds (with local LLMs) on average, making input characterization possible to be integrated into the workflow of program compilations. Optimizations on those identified input features show similar or even better results than those identified by previous profiling-based methods, leading to optimizations that yield 92.6% accuracy in selecting the appropriate adaptive OpenMP parallelization decisions, and 20-30% performance improvement of serverless computing while reducing resource usage by 50-60%.

Recent years have witnessed increasing interest in machine learning inferences on serverless computing for its auto-scaling and cost effective properties. Existing serverless computing, however, lacks effective job scheduling methods to handle the schedule space dramatically expanded by GPU sharing, task batching, and intertask relations. Prior solutions have dodged the issue by neglecting some important factors, leaving some large performance potential locked. This paper presents ESG, a new scheduling algorithm that directly addresses the difficulties. ESG treats sharable GPU as a first-order factor in scheduling. It employs an optimality-guided adaptive method by combining A*-search and a novel dual-blade pruning to dramatically prune the scheduling space without compromising the quality. It further introduces a novel method, dominator-based SLO distribution, to ensure the scalability of the scheduler. The results show that ESG can significantly improve the SLO hit rates (61%-80%) while saving 47%-187% costs over prior work.

Deterministic MultiThreading (DMT) systems eliminate nondeterminism from the dynamic executions of multithreaded programs. They can greatly simplify multithreaded programming and ease the deployment of systems that rely on replication. We first categorize and compare existing DMT system designs along three axes, incorporating the most recent advances in DMT systems. From our study, we conclude that stable synchronization determinism is the most cost-effective design, and it is thus the focus of our work.

Server-side web applications are vulnerable to request races. While some previous studies of real-world request races exist, they primarily focus on the root cause of these bugs. To better combat request races in server-side web applications, we need a deep understanding of their characteristics. In this paper, we provide a complementary focus on race effects and fixes with an enlarged set of request races from web applications developed with Object-Relational Mapping (ORM) frameworks. We revisit characterization questions used in previous studies on newly included request races, distinguish the external and internal effects of request races, and relate request-race fixes with concurrency control mechanisms in languages and frameworks for developing server-side web applications.

Material screening entails a large number of electronic structure simulations. Traditionally, these simulation runs are treated separately as solving independent Kohn-Sham (KS) equations. In this paper, we formulate material screening as an inter-job scheduling problem for solving a system of KS equations, and in doing so allowing one to explore different scheduling methods that use the results of some equations to expedite the solution of others. We propose the concept of sharing iterative simulation and employ several optimization methods to initialize a simulation run using the distribution of particles from similar jobs as the initial condition. More specifically, we propose two similarity metrics, one qualitative and the other quantitative, to predict the simulation runtime of a material screen job based on its similarity to other jobs. Accordingly, we present two inter-job scheduling algorithms that make use the qualitative and quantitative similarity information. We conducted extensive experiments on the Sunway TaihuLight supercomputer for a practical material screening problem to evaluate the performance of the two scheduling algorithms using the proposed similarity metrics. We show that the total time required to run the large number of material screening jobs can be significantly reduced, and the algorithms are robust even with moderate inaccurate prediction on the simulation runtime. The quantitative algorithm achieves better results than the qualitative algorithm using more accurate prediction and thus achieving more significant runtime reduction.