Device Hopping

Existing OS techniques for homogeneous many-core systems make it simple for single and multithreaded applications to migrate between cores. Heterogeneous systems do not benefit so fully from this flexibility, and applications that cannot migrate in mid-execution may lose potential performance. The situation is particularly challenging when a switch of language runtime would be desirable in conjunction with a migration. We present a case study in making heterogeneous CPU + GPU systems more flexible in this respect. Our technique for fine-grained application migration, allows switches between OpenMP, OpenCL, and CUDA execution, in conjunction with migrations from GPU to CPU, and CPU to GPU. To achieve this, we subdivide iteration spaces into slices, and consider migration on a slice-by-slice basis. We show that slice sizes can be learned offline by machine learning models. To further improve performance, memory transfers are made migration-aware. The complexity of the migration capability is hidden from programmers behind a high-level programming model. We present a detailed evaluation of our mid-kernel migration mechanism with the First Come, First Served scheduling policy. We compare our technique in a focused evaluation scenario against idealized kernel-by-kernel scheduling, which is typical for current systems, and makes perfect kernel to device scheduling decisions, but cannot migrate kernels mid-execution. Models show that up to 1.33× speedup can be achieved over these systems by adding fine-grained migration. Our experimental results with all nine applicable SHOC and Rodinia benchmarks achieve speedups of up to 1.30× (1.08× on average) over an implementation of a perfect but kernel-migration incapable scheduler when migrated to a faster device. Our mechanism and slice size choices introduce an average slowdown of only 2.44% if kernels never migrate. Lastly, our programming model reduces the code size by at least 88% if compared to manual implementations of migratable kernels.

Optimal Scheduling of Proactive Service with Customer Deterioration and Improvement

Service systems are typically limited resource environments where scarce capacity is reserved for the most urgent customers. However, there has been a growing interest in the use of proactive service when a less urgent customer may become urgent while waiting. On one hand, providing service for customers when they are less urgent could mean that fewer resources are needed to fulfill their service requirement. On the other hand, using limited capacity for customers who may never need the service in the future takes the capacity away from other more urgent customers who need it now. To understand this tension, we propose a multiserver queueing model with two customer classes: moderate and urgent. We allow customers to transition classes while waiting. In this setting, we characterize how moderate and urgent customers should be prioritized for service when proactive service for moderate customers is an option. We identify an index, the modified [Formula: see text]-index, which plays an important role in determining the optimal scheduling policy. This index lends itself to an intuitive interpretation of how to balance holding costs, service times, abandonments, and transitions between customer classes. This paper was accepted by David Simchi-Levi, stochastic models and simulation.

Analyzing Isolation in Mobile Systems

As a safety-critical issue in complex mobile systems, isolation requires two or more mobile objects not to appear in the same place simultaneously. To ensure such isolation, a scheduling policy is needed to control and coordinate the movement of mobile objects. Unfortunately, existing task scheduling theories fails in providing effective solutions, because it is hardly possible to decompose a complex mobile system into multiple independent tasks. To solve this problem, a more fine-grained event scheduling is proposed in this paper to generate scheduling policies which can ensure the isolation of mobile objects. After defining event scheduling based on event-based formal models called dependency structures, a new event scheduling theory for mobile systems is developed accordingly. Then an algorithm for generating an event scheduling policy is proposed to implement the required isolation. Simulation experiments are conducted to prove the result of our theoretical analysis and show the effectiveness and scalability of the approach.

Incentives for Shared Services: Multiserver Queueing Systems with Priorities

Problem definition: We study shared service whereby multiple independent service providers collaborate by pooling their resources into a shared service center (SSC). The SSC deploys an optimal priority scheduling policy for their customers collectively by accounting for their individual waiting costs and service-level requirements. We model the SSC as a multiclass [Formula: see text] queueing system subject to service-level constraints. Academic/practical relevance: Shared services are increasingly popular among firms for saving operational costs and improving service quality. One key issue in fostering collaboration is the allocation of costs among different firms. Methodology: To incentivize collaboration, we investigate cost allocation rules for the SSC by applying concepts from cooperative game theory. Results: To empower our analysis, we show that a cooperative game with polymatroid optimization can be analyzed via simple auxiliary games. By exploiting the polymatroidal structures of the multiclass queueing systems, we show when the games possess a core allocation. We explore the extent to which our results remain valid for some general cases. Managerial implications: We provide operational insights and guidelines on how to allocate costs for the SSC under the multiserver queueing context with priorities.

Control of Fork-Join Processing Networks with Multiple Job Types and Parallel Shared Resources

A fork-join processing network is a queueing network in which tasks associated with a job can be processed simultaneously. Fork-join processing networks are prevalent in computer systems, healthcare, manufacturing, project management, justice systems, and so on. Unlike the conventional queueing networks, fork-join processing networks have synchronization constraints that arise because of the parallel processing of tasks and can cause significant job delays. We study scheduling in fork-join processing networks with multiple job types and parallel shared resources. Jobs arriving in the system fork into arbitrary number of tasks, then those tasks are processed in parallel, and then they join and leave the network. There are shared resources processing multiple job types. We study the scheduling problem for those shared resources (i.e., which type of job to prioritize at any given time) and propose an asymptotically optimal scheduling policy in diffusion scale.

SRPT Scheduling Discipline in Many-Server Queues with Impatient Customers

The shortest-remaining-processing-time (SRPT) scheduling policy has been extensively studied, for more than 50 years, in single-server queues with infinitely patient jobs. Yet, much less is known about its performance in multiserver queues. In this paper, we present the first theoretical analysis of SRPT in multiserver queues with abandonment. In particular, we consider the [Formula: see text] queue and demonstrate that, in the many-sever overloaded regime, performance in the SRPT queue is equivalent, asymptotically in steady state, to a preemptive two-class priority queue where customers with short service times (below a threshold) are served without wait, and customers with long service times (above a threshold) eventually abandon without service. We prove that the SRPT discipline maximizes, asymptotically, the system throughput, among all scheduling disciplines. We also compare the performance of the SRPT policy to blind policies and study the effects of the patience-time and service-time distributions. This paper was accepted by Baris Ata, stochastic models & simulation.

Globally Scheduling Volunteer Computing

Volunteer computing uses millions of consumer computing devices (desktop and laptop computers, tablets, phones, appliances, and cars) to do high-throughput scientific computing. It can provide Exa-scale capacity, and it is a scalable and sustainable alternative to data-center computing. Currently, about 30 science projects use volunteer computing in areas ranging from biomedicine to cosmology. Each project has application programs with particular hardware and software requirements (memory, GPUs, VM support, and so on). Each volunteered device has specific hardware and software capabilities, and each device owner has preferences for which science areas they want to support. This leads to a scheduling problem: how to dynamically assign devices to projects in a way that satisfies various constraints and that balances various goals. We describe the scheduling policy used in Science United, a global manager for volunteer computing.

A UoI-Optimal Policy for Timely Status Updates with Resource Constraint

Timely status updates are critical in remote control systems such as autonomous driving and the industrial Internet of Things, where timeliness requirements are usually context dependent. Accordingly, the Urgency of Information (UoI) has been proposed beyond the well-known Age of Information (AoI) by further including context-aware weights which indicate whether the monitored process is in an emergency. However, the optimal updating and scheduling strategies in terms of UoI remain open. In this paper, we propose a UoI-optimal updating policy for timely status information with resource constraint. We first formulate the problem in a constrained Markov decision process and prove that the UoI-optimal policy has a threshold structure. When the context-aware weights are known, we propose a numerical method based on linear programming. When the weights are unknown, we further design a reinforcement learning (RL)-based scheduling policy. The simulation reveals that the threshold of the UoI-optimal policy increases as the resource constraint tightens. In addition, the UoI-optimal policy outperforms the AoI-optimal policy in terms of average squared estimation error, and the proposed RL-based updating policy achieves a near-optimal performance without the advanced knowledge of the system model.