Search for: All records

With the increasing popularity of large deep learning model serving workloads, there is a pressing need to reduce the energy consumption of a model-serving cluster while maintaining satisfied throughput or model-serving latency requirements. Model multiplexing approaches such as model parallelism, model placement, replication, and batching aim to optimize the model-serving performance. However, they fall short of leveraging the GPU frequency scaling opportunity for power saving. In this paper, we demonstrate (1) the benefits of GPU frequency scaling in power saving for model serving; and (2) the necessity for co-design and optimization of fine grained model multiplexing and GPU frequency scaling. We explore the co-design space and present a novel power-aware model-serving system, μ-Serve. μ-Serve is a model-serving framework that optimizes the power consumption and model serving latency/throughput of serving multiple ML models efficiently in a homogeneous GPU cluster. Evaluation results on production workloads show that μ-Serve achieves 1.2–2.6× power saving by dynamic GPU frequency scaling (up to 61% reduction) without SLO attainment violations. 

The emergence of ML in various cloud system management tasks (e.g., workload autoscaling and job scheduling) has become a core driver of ML-centric cloud platforms. However, there are still numerous algorithmic and systems challenges that prevent ML-centric cloud platforms from being production-ready. In this paper, we focus on the challenges of model performance variability and costly model retraining, introduced by dynamic workload patterns and heterogeneous applications and infrastructures in cloud environments. To address these challenges, we present FLASH, an extensible framework for fast model adaptation in ML-based system management tasks. We show how FLASH leverages existing ML agents and their training data to learn to generalize across applications/environments with meta-learning. FLASH can be easily integrated with an existing ML-based system management agent with a unified API. We demonstrate the use of FLASH by implementing three existing ML agents that manage (1) resource configurations, (2) autoscaling, and (3) server power. Our experiments show that FLASH enables fast adaptation to new, previously unseen applications/environments (e.g., 5.5× faster than transfer learning in the autoscaling task), indicating significant potential for adopting ML-centric cloud platforms in production. 

The emergence of ML in various cloud system management tasks (e.g., workload autoscaling and job scheduling) has become a core driver of ML-centric cloud platforms. However, there are still numerous algorithmic and systems challenges that prevent ML-centric cloud platforms from being production-ready. In this paper, we focus on the challenges of model performance variability and costly model retraining, introduced by dynamic workload patterns and heterogeneous applications and infrastructures in cloud environments. To address these challenges, we present FLASH, an extensible framework for fast model adaptation in ML-based system management tasks. We show how FLASH leverages existing ML agents and their training data to learn to generalize across applications/environments with meta-learning. FLASH can be easily integrated with an existing ML-based system management agent with a unified API. We demonstrate the use of FLASH by implementing three existing ML agents that manage (1) resource configurations, (2) autoscaling, and (3) server power. Our experiments show that FLASH enables fast adaptation to new, previously unseen applications/environments (e.g., 5.5× faster than transfer learning in the autoscaling task), indicating significant potential for adopting ML-centric cloud platforms in production. 

With the increasing popularity of large deep learning model-serving workloads, there is a pressing need to reduce the energy consumption of a model-serving cluster while maintaining satisfied throughput or model-serving latency requirements. Model multiplexing approaches such as model parallelism, model placement, replication, and batching aim to optimize the model-serving performance. However, they fall short of leveraging the GPU frequency scaling opportunity for power saving. In this paper, we demonstrate (1) the benefits of GPU frequency scaling in power saving for model serving; and (2) the necessity for co-design and optimization of fine-grained model multiplexing and GPU frequency scaling. We explore the co-design space and present a novel power-aware model-serving system, μ-Serve. μ-Serve is a model-serving framework that optimizes the power consumption and model-serving latency/throughput of serving multiple ML models efficiently in a homogeneous GPU cluster. Evaluation results on production workloads show that μ-Serve achieves 1.2–2.6× power saving by dynamic GPU frequency scaling (up to 61% reduction) without SLO attainment violations. 

This paper addresses the challenge of ensuring the safety of autonomous vehicles (AVs, also called ego actors) in realworld scenarios where AVs are constantly interacting with other actors. To address this challenge, we introduce iPrism which incorporates a new risk metric – the Safety-Threat Indicator (STI). Inspired by how experienced human drivers proactively mitigate hazardous situations, STI quantifies actor-related risks by measuring the changes in escape routes available to the ego actor. To actively mitigate the risk quantified by STI and avert accidents, iPrism also incorporates a reinforcement learning (RL) algorithm (referred to as the Safety-hazard Mitigation Controller (SMC)) that learns and implements optimal risk mitigation policies. Our evaluation of the success of the SMC is based on over 4800 NHTSA-based safety-critical scenarios. The results show that (i) STI provides up to 4.9× longer lead-time for-mitigating-accidents compared to widely-used safety and planner-centric metrics, (ii) SMC significantly reduces accidents by 37% to 98% compared to a baseline Learning-by-Cheating (LBC) agent, and (iii) in comparison with available state-of-the-art safety hazard mitigation agents, SMC prevents up to 72.7% of accidents that the selected agents are unable to avoid. All code, model weights, and evaluation scenarios and pipelines used in this paper are available at: https://zenodo.org/doi/10.5281/ zenodo.10279653. 

In this paper, we i) analyze and classify real-world failures of Kubernetes (the most popular container orchestration system), ii) develop a framework to perform a fault/error injection campaign targeting the data store preserving the cluster state, and iii) compare results of our fault/error injection experiments with real-world failures, showing that our fault/error injections can recreate many real-world failure patterns. The paper aims to address the lack of studies on systematic analyses of Kubernetes failures to date. Our results show that even a single fault/error (e.g., a bit-flip) in the data stored can propagate, causing cluster-wide failures (3% of injections), service networking issues (4%), and service under/over provisioning (24%). Errors in the fields tracking dependencies between object caused 51% of such cluster-wide failures. We argue that controlled fault/error injection-based testing should be employed to proactively assess Kubernetes' resiliency and guide the design of failure mitigation strategies. 

Multi-agent reinforcement learning (MARL) has primarily focused on solving a single task in isolation, while in practice the environment is often evolving, leaving many related tasks to be solved. In this paper, we investigate the benefits of meta-learning in solving multiple MARL tasks collectively. We establish the first line of theoretical results for meta-learning in a wide range of fundamental MARL settings, including learning Nash equilibria in two-player zero-sum Markov games and Markov potential games, as well as learning coarse correlated equilibria in general-sum Markov games. Under natural notions of task similarity, we show that meta-learning achieves provable sharper convergence to various game-theoretical solution concepts than learning each task separately. As an important intermediate step, we develop multiple MARL algorithms with initialization-dependent convergence guarantees. Such algorithms integrate optimistic policy mirror descents with stage-based value updates, and their refined convergence guarantees (nearly) recover the best known results even when a good initialization is unknown. To our best knowledge, such results are also new and might be of independent interest. We further provide numerical simulations to corroborate our theoretical findings. 

This paper addresses the urgent need to transition to global net-zero carbon emissions by 2050 while retaining the ability to meet joint performance and resilience objectives. The focus is on the computing infrastructures, such as hyperscale cloud datacenters, that consume significant power, thus producing increasing amounts of carbon emissions. Our goal is to (1) optimize the usage of green energy sources (e.g., solar energy), which is desirable but expensive and relatively unstable, and (2) continuously reduce the use of fossil fuels, which have a lower cost but a significant negative societal impact. Meanwhile, cloud datacenters strive to meet their customers’ requirements, e.g., service-level objectives (SLOs) in application latency or throughput, which are impacted by infrastructure resilience and availability. We propose a scalable formulation that combines sustainability, cloud resilience, and performance as a joint optimization problem with multiple interdependent objectives to address these issues holistically. Given the complexity and dynamicity of the problem, machine learning (ML) approaches, such as reinforcement learning, are essential for achieving continuous optimization. Our study highlights the challenges of green energy instability which necessitates innovative MLcentric solutions across heterogeneous infrastructures to manage the transition towards green computing. Underlying the MLcentric solutions must be methods to combine classic system resilience techniques with innovations in real-time ML resilience (not addressed heretofore). We believe that this approach will not only set a new direction in the resilient, SLO-driven adoption of green energy but also enable us to manage future sustainable systems in ways that were not possible before. 

Workload autoscaling is widely used in public and private cloud systems to maintain stable service performance and save resources. However, it remains challenging to set the optimal resource limits and dynamically scale each workload at runtime. Reinforcement learning (RL) has recently been proposed and applied in various systems tasks, including resource management. In this paper, we first characterize the state-of-the-art RL approaches for workload autoscaling in a public cloud and point out that there is still a large gap in taking the RL advances to production systems. We then propose AWARE, an extensible framework for deploying and managing RL-based agents in production systems. AWARE leverages meta-learning and bootstrapping to (a) automatically and quickly adapt to different workloads, and (b) provide safe and robust RL exploration. AWARE provides a common OpenAI Gym-like RL interface to agent developers for easy integration with different systems tasks. We illustrate the use of AWARE in the case of workload autoscaling. Our experiments show that AWARE adapts a learned autoscaling policy to new workloads 5.5x faster than the existing transfer-learning-based approach and provides stable online policy-serving performance with less than 3.6% reward degradation. With bootstrapping, AWARE helps achieve 47.5% and 39.2% higher CPU and memory utilization while reducing SLO violations by a factor of 16.9x during policy training. 

Serverless Function-as-a-Service (FaaS) offers improved programmability for customers, yet it is not server-“less” and comes at the cost of more complex infrastructure management (e.g., resource provisioning and scheduling) for cloud providers. To maintain function service-level objectives (SLOs) and improve resource utilization efficiency, recent research has been focused on applying online learning algorithms such as reinforcement learning (RL) to manage resources. Compared to rule-based solutions with heuristics, RL-based approaches eliminate humans in the loop and avoid the painstaking generation of heuristics. Despite the initial success of applying RL, we first show in this paper that the state-of-the-art single-agent RL algorithm (S-RL) suffers up to 4.8x higher p99 function latency degradation on multi-tenant serverless FaaS platforms compared to isolated environments and is unable to converge during training. We then design and implement a scalable and incremental multi-agent RL framework based on Proximal Policy Optimization (SIMPPO). Our experiments on widely used serverless benchmarks demonstrate that in multi-tenant environments, SIMPPO enables each RL agent to efficiently converge during training and provides online function latency performance comparable to that of S-RL trained in isolation (which we refer to as the baseline for assessing RL performance) with minor degradation (<9.2%). In addition, SIMPPO reduces the p99 function latency by 4.5x compared to S-RL in multi-tenant cases.