To facilitate various Machine Learning (ML) training and inference tasks, enterprises tend to build large and expensive clusters and share them among different teams for diverse ML workloads. Virtualized platforms (containers/VMs) and schedulers are typically deployed to allow such access, manage heterogeneous resources and schedule ML jobs in these clusters. However, allocating resource budgets for different ML jobs to achieve best performance and cluster resource efficiency remains a significant challenge. This work proposes NEARCHUS to accelerate distributed ML training while ensuring high resource efficiency by using adaptive resource allocation. NEARCHUS automatically identifies potential performance bottlenecks for running jobs and re-allocates resources to provide optimized run-time performance with high resource efficiency. NEARCHUS's resource configuration significantly improves the training speed of individual jobs up to 71.4%-129.1% against state-of-the-art resource schedulers, and reduces job completion and queuing time by 35.6% and 67.8%, respectively.

This paper presents an IRB-approved human study to capture data to build models for human frustration prediction of computer users. First, an application was developed that ran in the user’s computer/laptop/VM with Linux 20.04. Then, the application collected a variety of data from their computers, including: mouse clicks, movements and scrolls; the pattern of keyboard keys clicks; user audio features; and head movements through the user video; System-wide information such as computation, memory usage, network bandwidth, and input/output bandwidth of the running applications in the computer and user frustrations. Finally, the application sent the data to the cloud. After two weeks of data collection, supervised and semi-supervised models were developed offline to predict user frustration with the computer using the collected data. A semi-supervised model using a generative adversarial network (GAN) resulted in the highest accuracy of 90%.

Today, video cameras are ubiquitously deployed. These cameras collect, stream, store, and analyze video footage for a variety of use cases, ranging from surveillance, retail analytics, architectural engineering, and more. At the same time, many citizens are becoming weary of the amount of personal data captured, along with the algorithms and datasets used to process video pipelines. This work investigates how users can opt-out of such pipelines by explicitly providing consent to be recorded. An ideal system should obfuscate or otherwise cleanse non-consenting user data, ideally before a user even enters the video processing pipeline itself. We present a system, called Consent-Box, that enables obfuscation of users without using complex or personally-identifying vision techniques. Instead, a user's location on a video frame is estimated via Wi-Fi localization of a user's mobile device. This estimation allows us to remove individuals from frames before those frames enter complex vision pipelines.

In the past, researchers designed, deployed, and evaluated Wi-Fi based localization techniques in order to locate users and devices without adding extra or costly infrastructure. However, as infrastructure deployments change, one must reexamine the role of Wi-Fi localization. Today, cameras are becoming increasingly deployed, and therefore this work examines how contextual and vision data obtained from cameras can be integrated with Wi-Fi localization techniques. We present an approach called CALM that works on commodity APs and cameras. Our approach contains several contributions: a camera line fitting technique to restrict the search space of candidate locations, single AP and camera localization via a deprojection scheme inspired from 3D cameras, simple and robust AP weighting that analyzes the context of users via the camera, and a new virtual camera methodology to scale analysis. We motivate our scheme by analyzing real camera and AP topologies from a major vendor. Our evaluation over 9 rooms and 102,300 wireless readings shows CALM can obtain decimeter-level accuracy, improving performance over previous Wi-Fi techniques like FTM by $2.7\times$ and SpotFi by $2.3\times$.

This paper pushes the limits of automated resource allocation in container environments. Recent works set container CPU and memory limits by automatically scaling containers based on past resource usage. However, these systems are heavyweight and run on coarse-grained time scales, resulting in poor performance when predictions are incorrect. We propose Escra, a container orchestrator that enables fine-grained, event-based resource allocation for a single container and distributed resource allocation to manage a collection of containers. Escra performs resource allocation on sub-second intervals within and across hosts, allowing operators to cost-effectively scale resources without performance penalty. We evaluate Escra on two types of containerized applications: microservices and serverless functions. In microservice environments, fine-grained and event-based resource allocation can reduce application latency by up to 96.9% and increase throughput by up to 3.2x when compared against the current state-of-the-art. Escra can increase performance while simultaneously reducing 50th and 99th%ile CPU waste by over 10x and 3.2x, respectively. In serverless environments, Escra can reduce CPU reservations by over 2.1x and memory reservations by more than 2x while maintaining similar end-to-end performance.

Today's operating systems typically apply a one-size-fits-all approach to resource management, such as applying a scheduler that treats all processes of equal importance. The goal of this paper is to explore a learning-based approach to resource management in modern operating systems in which the OS automatically learns what tasks the user deems to be most important at that time and seamlessly adjusts allocation of CPU, memory, I/O, and network bandwidth to prioritize user preferences on demand. We demonstrate an implementation of such a learning-based OS in Linux and present evaluation results showing that a reinforcement learning-based approach can rapidly learn and adjust system resources to meet user demands.

Serverless Computing is a new cloud computing paradigm wherein people in academia and industry are actively proposing either interesting improvements or building excellent applications on top of it. AWS, Google Cloud, Microsoft Azure, and IBM are popular samples of public clouds that offer Function-as-a-Service on top of their Serverless Computing platforms. Although this paradigm has had numerous advantages for software developers and programmers, it has introduced new challenges to cloud providers. Factors like fine-grained pricing and pay-as-you-go manner, eliminating the responsibility of resource management on the developer side, promises of elasticity and highly-available service, fault tolerance, auto-scaling, and being able to run embarrassingly parallel jobs make it a suitable platform for developers. On the other hand, efficient resource management, offering low-latency service, and providing proper security/isolation have been partly the main challenges introduced on the cloud provider side. This paper presents a literature review on today's Serverless platform optimizations and extensions that people have proposed and implemented to further capitalize the Serverless infras-tructure. In the end, we will provide the current Serverless paradigm's limitations and a few future directions and research opportunities regarding Serverless Computing.

Serverless computing is a rapidly growing paradigm that easily harnesses the power of the cloud. With serverless computing, developers simply provide an event-driven function to cloud providers, and the provider seamlessly scales function invocations to meet demands as event-triggers occur. As current and future serverless offerings support a wide variety of serverless applications, effective techniques to manage serverless workloads becomes an important issue. This work examines current management and scheduling practices in cloud providers, uncovering many issues including inflated application run times, function drops, inefficient allocations, and other undocumented and unexpected behavior. To fix these issues, a new quality-of-service function scheduling and allocation framework, called Sequoia, is designed. Sequoia allows developers or administrators to easily def ne how serverless functions and applications should be deployed, capped, prioritized, or altered based on easily configured, flexible policies. Results with controlled and realistic workloads show Sequoia seamlessly adapts to policies, eliminates mid-chain drops, reduces queuing times by up to 6.4X, enforces tight chain-level fairness, and improves run-time performance up to 25X.

This paper presents a new approach to memory disaggregation called FluidMem that leverages the userfault mechanism in Linux to achieve full memory disaggregation in software. FluidMem enables dynamic and transparent resizing of an unmodified Virtual Machine's (VM's) memory footprint in the cloud. As a result, a VM's memory footprint can seamlessly scale over multiple machines or even be downsized to a near-zero footprint on a given server. FluidMem's architecture provides flexibility to cloud operators to manage remote memory without requiring guest intervention, while also supporting paging out the entirety of a VM's pages within its address space. FluidMem integrates with a remote memory backend in a modular way, easily supporting systems such as RAMCloud to harness remote memory. We demonstrate FluidMem outperforms an existing memory disaggregation approach based on network swap. Microbenchmarks are evaluated to characterize the latency of different components of the FluidMem architecture, and two memory-intensive applications are demonstrated using FluidMem, the Graph500 benchmark, and MongoDB. Additionally, we show FluidMem can flexibly and efficiently grow and shrink the memory footprint of a VM as defined by a cloud provider.

This paper introduces a new primitive to serverless language runtimes called freshen. With freshen, developers or providers specify functionality to perform before a given function executes. This proactive technique allows for overheads associated with serverless functions to be mitigated at execution time, which improves function responsiveness. We show various predictive opportunities exist to run freshen within reasonable time windows. A high-level design and implementation are described, along with preliminary results to show the potential benefits of our scheme.

Wireless network management is important to ensure the performance, utilization, allocation, and robustness of the network is optimized. Until now, wireless network management has typically been dictated by in-band information, such as wireless measurements, client locations, or even device state. This position paper explores fundamental new ways to manage the network by utilizing out-of-band data provided by a rich deployment of sensors. Out-of-band data can capture information about the users, objects, or environments associated with a network device, meaning that richer contextual policies can be implemented in the network. We propose an architecture called SenseNet, which builds upon three recent trends: (1) the massive deployment of sensors today, (2) the existence of deep-learning algorithms to glean meaningful insights from the sensory data, and (3) the provisioning of edge computing resources to provide real-time actuation of new sensor-based policies. A brief evaluation shows the feasibility and motivates SenseNet and afterwards challenges towards practical deployment are discussed.

Serverless computing is a rapidly growing paradigm that easily harnesses the power of the cloud. With serverless computing, developers simply provide an event-driven function to cloud providers, and the provider seamlessly scales function invocations to meet demands as event-triggers occur. As current and future serverless offerings support a wide variety of serverless applications, effective techniques to manage serverless workloads becomes an important issue. This work examines current management and scheduling practices in cloud providers, uncovering many issues including inflated application run times, function drops, inefficient allocations, and other undocumented and unexpected behavior. To fix these issues, a new quality-of-service function scheduling and allocation framework, called Sequoia, is designed. Sequoia allows developers or administrators to easily def ne how serverless functions and applications should be deployed, capped, prioritized, or altered based on easily configured, flexible policies. Results with controlled and realistic workloads show Sequoia seamlessly adapts to policies, eliminates mid-chain drops, reduces queuing times by up to 6.4X, enforces tight chain-level fairness, and improves run-time performance up to 25X.

Efficient resource management in cloud computing research is a crucial problem because resource over-provisioning increases costs for cloud providers and cloud customers; resource under-provisioning increases the application latency, and it may violate service level agreements, which eventually makes cloud providers lose their customers and income. As a result, researchers have been striving to develop optimal resource management in cloud computing environments in different ways, such as container placement, job scheduling and multi-resource scheduling. Machine learning techniques are extensively used in this area. In this paper, we present a comprehensive survey on the projects that leveraged machine learning techniques for resource management solutions in the cloud computing environment. At the end, we provide a comparison between these projects. Furthermore, we propose some future directions that will guide researchers to advance this field.

Containers are a popular mechanism used among application developers when deploying their systems on cloud platforms. Both developers and cloud providers are constantly looking to simplify container management, provisioning, and monitoring. In this paper, we present a container management layer that sits beside a container orchestrator that runs, what we call, Elastic Containers. Each elastic container contains multiple subcontainers that are connected to a centralized Global Cloud Manager (GCM). The GCM gathers subcontainer resource utilization information directly from inside each kernel running the subcontainers. The GCM then tries to efficiently and optimally distribute resources between the application subcontainers residing on a distributed environment.

As personal interactive devices become more ingrained into our daily lives, it becomes more important to understand how seamless interaction with those devices can be fostered. A typical mechanism to interface with a personal device is via a touch screen, in which users use their fingertip or stylus to scroll, type, select, or otherwise control device usage. Touch-based techniques, however, can become restrictive or inconvenient under a variety of scenarios. For example, personal devices such as phones or tablets are continuously increasing in size, making one-handed interaction difficult because one cannot easily hold the phone and touch the screen (with the thumb) at the same time with one hand. Therefore, in this demo, we present a new technique to interact with personal devices in which the screen and touch screen interactions can adapt to a user's grip or current touch constraints.

Containers are quickly increasing in popularity as the mechanism to deploy computation in the cloud. In order to provide consistent and reliable performance, cloud providers must ensure containers cannot adversely interfere with one another. Because containers share the same underlying OS, it is more challenging to provide isolation in a container-based framework than a traditional VM-based framework. And while many schemes can isolate CPU, memory, disk, or network bandwidth in multitenant environments, less attention has been paid to how the time spent processing network traffic affects isolation on the host server. This paper shows computational overhead associated with the network stack can break isolation in container-based environments. Specifically, a container with heavy network traffic can decrease the computation available to other containers sharing the same server. We propose a scheme, called Iron, that accounts for the time spent in the networking stack on behalf of a container and ensures this processing cannot adversely impact colocated containers through novel enforcement mechanisms. Our results show Iron effectively provides isolation under realistic and adversarial conditions, limiting interference-based slowdowns as high as 6× to less than 5%.

In serverless computing, developers define a function to handle an event, and the serverless framework horizontally scales the application as needed. The downside of this function-based abstraction is it limits the type of application supported and places a bound on the function to be within the physical resource limitations of the server the function executes on. In this paper we propose a new abstraction for serverless computing: a developer supplies a process and the serverless framework seamlessly scales out the process's resource usage across the datacenter. This abstraction enables processing to not only be more general purpose, but also allows a process to break out of the limitations of a single server -making serverless computing more serverless. To realize this abstraction, we propose ServerlessOS, comprised of three key components: (i) a new disaggregation model, which leverages disaggregation for abstraction, but enables resources to move fluidly between servers for performance; (ii) a cloud orchestration layer which manages fine-grained resource allocation and placement throughout the application's lifetime via local and global decision making; and (iii) an isolation capability that enforces data and resource isolation across disaggregation, effectively extending Linux cgroup functionality to span servers.

1 I o h m f a k s h L p th c I p r c in 2 F a c f s F Dem Everyd Chungkuk Y KAIST 1. INTROD Imagine you c objects using hardware. To t mobile platform fingertip locali as TV screen, kitchen counte surficial intera hardware not c LED, pair of position on a su he 3-D localiz camera of a com In this demo, principle of ref reflection caste challenges. Al nteractive appl