Search

Cloud systems are the backbone of today's computing industry. Yet, these systems remain complicated to design, build, operate, and improve. All these tasks require significant manual effort by both developers and operators of these systems. To reduce this manual burden, in this paper we set forth a vision for achieving holistic automation, intent-based system design and operation. We propose intent as a new abstraction within the context of system design and operation. Intent encodes the functional and operational requirements of the system at a high-level, which can be used to automate design, implementation, operation, and evolution of systems. We detail our vision of intent-based system design, highlight its four key components, and provide a roadmap for the community to enable autonomous systems.

The rapid increase in LLM ubiquity and scale levies unprecedented demands on computing infrastructure. These demands not only incur large compute and memory resources, but also significant energy, yielding large operational and embodied carbon emissions. In this work, we present two main observations. First, while GPUs dominate operational carbon, host processing systems (e.g., CPUs, memory, storage) dominate embodied carbon. Second, based on traces from production deployment of two Generative AI services in the cloud, offline, batch-inference accounts for a significant portion (up to 55\%) of serving capacity. We propose four pillars of carbon-conscious infrastructure design for LLM serving systems: \textbf{\textit{Reduce, Reuse, Rightsize, and Recycle}}. We demonstrate that EcoServe can lower carbon emissions by up to 47\%, compared to performance, energy, and cost-optimized design points, while maintaining performance targets and SLOs.

Recent advances in generative AI have led to large multi-modal models (LMMs) capable of simultaneously processing inputs of various modalities such as text, images, video, and audio. While these models demonstrate impressive capabilities, efficiently serving them in production environments poses significant challenges due to their complex architectures and heterogeneous resource requirements. We present the first comprehensive systems analysis of two prominent LMM architectures, decoder-only and cross-attention, on six representative open-source models. We investigate their multi-stage inference pipelines and resource utilization patterns that lead to unique systems design implications. We also present an in-depth analysis of production LMM inference traces, uncovering unique workload characteristics, including variable, heavy-tailed request distributions, diverse modal combinations, and bursty traffic patterns. Our key findings reveal that different LMM inference stages exhibit highly heterogeneous performance characteristics and resource demands, while concurrent requests across modalities lead to significant performance interference. To address these challenges, we propose a decoupled serving architecture that enables independent resource allocation and adaptive scaling for each stage. We further propose optimizations such as stage colocation to maximize throughput and resource utilization while meeting the latency objectives.

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.

Compute Express Link (CXL) is widely-supported interconnect standard that promises to enable memory disaggregation in data centers. CXL allows for memory pooling, which can be used to create a shared memory space across multiple servers. However, CXL does not specify how to actually build a memory pool. Existing proposals for CXL memory pools are expensive, as they require CXL switches or large multi-headed devices. In this paper, we propose a new design for CXL memory pools that is cost-effective. We call these designs Octopus topologies. Our design uses small CXL devices that can be made cheaply and offer fast access latencies. Specifically, we propose asymmetric CXL topologies where hosts connect to different sets of CXL devices. This enables pooling and sharing memory across multiple hosts even as each individual CXL device is only connected to a small number of hosts. Importantly, this uses hardware that is readily available today. We also show the trade-off in terms of CXL pod size and cost overhead per host. Octopus improves the Pareto frontier defined by prior policies, e.g., offering to connect 3x as many hosts at 17% lower cost per host.

The rising demand for generative large language models (LLMs) poses challenges for thermal and power management in cloud datacenters. Traditional techniques often are inadequate for LLM inference due to the fine-grained, millisecond-scale execution phases, each with distinct performance, thermal, and power profiles. Additionally, LLM inference workloads are sensitive to various configuration parameters (e.g., model parallelism, size, and quantization) that involve trade-offs between performance, temperature, power, and output quality. Moreover, clouds often co-locate SaaS and IaaS workloads, each with different levels of visibility and flexibility. We propose TAPAS, a thermal- and power-aware framework designed for LLM inference clusters in the cloud. TAPAS enhances cooling and power oversubscription capabilities, reducing the total cost of ownership (TCO) while effectively handling emergencies (e.g., cooling and power failures). The system leverages historical temperature and power data, along with the adaptability of SaaS workloads, to: (1) efficiently place new GPU workload VMs within cooling and power constraints, (2) route LLM inference requests across SaaS VMs, and (3) reconfigure SaaS VMs to manage load spikes and emergency situations. Our evaluation on a large GPU cluster demonstrates significant reductions in thermal and power throttling events, boosting system efficiency.

The demand for large language model (LLM) inference is gradually dominating the artificial intelligence workloads. Therefore, there is an urgent need for cost-efficient inference serving. Existing work focuses on single-worker optimization and lacks consideration of cluster-level management for both inference queries and computing resources. However, placing requests and managing resources without considering the query features easily causes SLO violations or resource underutilization. Providers are forced to allocate extra computing resources to guarantee user experience, leading to additional serving costs. In this paper we introduce Aladdin, a scheduler that co-adaptively places queries and scales computing resources with SLO awareness. For a stream of inference queries, Aladdin first predicts minimal computing resources and the corresponding serving workers' configuration required to fulfill the SLOs for all queries. Then, it places the queries to each serving worker according to the prefill and decode latency models of batched LLM inference to maximize each worker's utilization. Results show that Aladdin reduces the serving cost of a single model by up to 71% for the same SLO level compared with the baselines, which can be millions of dollars per year.

Today, cloud workloads are essentially opaque to the cloud platform. Typically, the only information the platform receives is the virtual machine (VM) type and possibly a decoration to the type (e.g., the VM is evictable). Similarly, workloads receive little to no information from the platform; generally, workloads might receive telemetry from their VMs or exceptional signals (e.g., shortly before a VM is evicted). The narrow interface between workloads and platforms has several drawbacks: (1) a surge in VM types and decorations in public cloud platforms complicates customer selection; (2) essential workload characteristics (e.g., low availability requirements, high latency tolerance) are often unspecified, hindering platform customization for optimized resource usage and cost savings; and (3) workloads may be unaware of potential optimizations or lack sufficient time to react to platform events. In this paper, we propose a framework, called Workload Intelligence (WI), for dynamic bi-directional communication between cloud workloads and cloud platform. Via WI, workloads can programmatically adjust their key characteristics, requirements, and even dynamically adapt behaviors like VM priorities. In the other direction, WI allows the platform to programmatically inform workloads about upcoming events, opportunities for optimization, among other scenarios. Because of WI, the cloud platform can drastically simplify its offerings, reduce its costs without fear of violating any workload requirements, and reduce prices to its customers on average by 48.8%.

The growing complexity of cloud based software systems has resulted in incident management becoming an integral part of the software development lifecycle. Root cause analysis (RCA), a critical part of the incident management process, is a demanding task for on-call engineers, requiring deep domain knowledge and extensive experience with a team's specific services. Automation of RCA can result in significant savings of time, and ease the burden of incident management on on-call engineers. Recently, researchers have utilized Large Language Models (LLMs) to perform RCA, and have demonstrated promising results. However, these approaches are not able to dynamically collect additional diagnostic information such as incident related logs, metrics or databases, severely restricting their ability to diagnose root causes. In this work, we explore the use of LLM based agents for RCA to address this limitation. We present a thorough empirical evaluation of a ReAct agent equipped with retrieval tools, on an out-of-distribution dataset of production incidents collected at Microsoft. Results show that ReAct performs competitively with strong retrieval and reasoning baselines, but with highly increased factual accuracy. We then extend this evaluation by incorporating discussions associated with incident reports as additional inputs for the models, which surprisingly does not yield significant performance improvements. Lastly, we conduct a case study with a team at Microsoft to equip the ReAct agent with tools that give it access to external diagnostic services that are used by the team for manual RCA. Our results show how agents can overcome the limitations of prior work, and practical considerations for implementing such a system in practice.

This report explores the use of kernel-bypass networking in FaaS runtimes and demonstrates how using Junction, a novel kernel-bypass system, as the backend for executing components in faasd can enhance performance and isolation. Junction achieves this by reducing network and compute overheads and minimizing interactions with the host operating system. Junctiond, the integration of Junction with faasd, reduces median and P99 latency by 37.33% and 63.42%, respectively, and can handle 10 times more throughput while decreasing latency by 2x at the median and 3.5 times at the tail.

Major cloud providers have employed advanced AI-based solutions like large language models to aid humans in identifying the root causes of cloud incidents. Despite the growing prevalence of AI-driven assistants in the root cause analysis process, their effectiveness in assisting on-call engineers is constrained by low accuracy due to the intrinsic difficulty of the task, a propensity for LLM-based approaches to hallucinate, and difficulties in distinguishing these well-disguised hallucinations. To address this challenge, we propose to perform confidence estimation for the predictions to help on-call engineers make decisions on whether to adopt the model prediction. Considering the black-box nature of many LLM-based root cause predictors, fine-tuning or temperature-scaling-based approaches are inapplicable. We therefore design an innovative confidence estimation framework based on prompting retrieval-augmented large language models (LLMs) that demand a minimal amount of information from the root cause predictor. This approach consists of two scoring phases: the LLM-based confidence estimator first evaluates its confidence in making judgments in the face of the current incident that reflects its ``grounded-ness" level in reference data, then rates the root cause prediction based on historical references. An optimization step combines these two scores for a final confidence assignment. We show that our method is able to produce calibrated confidence estimates for predicted root causes, validate the usefulness of retrieved historical data and the prompting strategy as well as the generalizability across different root cause prediction models. Our study takes an important move towards reliably and effectively embedding LLMs into cloud incident management systems.

Current serverless offerings give users a limited degree of flexibility for configuring the resources allocated to their function invocations by either coupling memory and CPU resources together or providing no knobs at all. These configuration choices simplify resource allocation decisions on behalf of users, but at the same time, create deployments that are resource inefficient. In this paper, we take a principled approach to the problem of resource allocation for serverless functions, allowing this choice to be made in an automatic way that leads to the best combination of performance and cost. In particular, we systematically explore the opportunities that come with decoupling memory and CPU resource allocations and also enabling the use of different VM types. We find a rich trade-off space between performance and cost. The provider can use this in a number of ways: from exposing all these parameters to the user, to eliciting preferences for performance and cost from users, or by simply offering the same performance with lower cost. This flexibility can also enable the provider to optimize its resource utilization and enable a cost-effective service with predictable performance. Our results show that, by decoupling memory and CPU allocation, there is potential to have up to 40% lower execution cost than the preset coupled configurations that are the norm in current serverless offerings. Similarly, making the correct choice of VM instance type can provide up to 50% better execution time. Furthermore, we demonstrate that providers can utilize different instance types for the same functions to maximize resource utilization while providing performance within 10-20% of the best resource configuration for each respective function.

Function-as-a-Service (FaaS) has become an increasingly popular way for users to deploy their applications without the burden of managing the underlying infrastructure. However, existing FaaS platforms rely on remote storage to maintain state, limiting the set of applications that can be run efficiently. Recent caching work for FaaS platforms has tried to address this problem, but has fallen short: it disregards the widely different characteristics of FaaS applications, does not scale the cache based on data access patterns, or requires changes to applications. To address these limitations, we present Faa\$T, a transparent auto-scaling distributed cache for serverless applications. Each application gets its own Faa\$T cache. After a function executes and the application becomes inactive, the cache is unloaded from memory with the application. Upon reloading for the next invocation, Faa\$T pre-warms the cache with objects likely to be accessed. In addition to traditional compute-based scaling, Faa\$T scales based on working set and object sizes to manage cache space and I/O bandwidth. We motivate our design with a comprehensive study of data access patterns in a large-scale commercial FaaS provider. We implement Faa\$T for the provider's production FaaS platform. Our experiments show that Faa\$T can improve performance by up to 92% (57% on average) for challenging applications, and reduce cost for most users compared to state-of-the-art caching systems, i.e. the cost of having to stand up additional serverful resources., Comment: 18 pages, 15 figures

High dimensional black-box optimization has broad applications but remains a challenging problem to solve. Given a set of samples $\{\vx_i, y_i\}$, building a global model (like Bayesian Optimization (BO)) suffers from the curse of dimensionality in the high-dimensional search space, while a greedy search may lead to sub-optimality. By recursively splitting the search space into regions with high/low function values, recent works like LaNAS shows good performance in Neural Architecture Search (NAS), reducing the sample complexity empirically. In this paper, we coin LA-MCTS that extends LaNAS to other domains. Unlike previous approaches, LA-MCTS learns the partition of the search space using a few samples and their function values in an online fashion. While LaNAS uses linear partition and performs uniform sampling in each region, our LA-MCTS adopts a nonlinear decision boundary and learns a local model to pick good candidates. If the nonlinear partition function and the local model fits well with ground-truth black-box function, then good partitions and candidates can be reached with much fewer samples. LA-MCTS serves as a \emph{meta-algorithm} by using existing black-box optimizers (e.g., BO, TuRBO) as its local models, achieving strong performance in general black-box optimization and reinforcement learning benchmarks, in particular for high-dimensional problems.

Efficient evaluation of a network architecture drawn from a large search space remains a key challenge in Neural Architecture Search (NAS). Vanilla NAS evaluates each architecture by training from scratch, which gives the true performance but is extremely time-consuming. Recently, one-shot NAS substantially reduces the computation cost by training only one supernetwork, a.k.a. supernet, to approximate the performance of every architecture in the search space via weight-sharing. However, the performance estimation can be very inaccurate due to the co-adaption among operations. In this paper, we propose few-shot NAS that uses multiple supernetworks, called sub-supernet, each covering different regions of the search space to alleviate the undesired co-adaption. Compared to one-shot NAS, few-shot NAS improves the accuracy of architecture evaluation with a small increase of evaluation cost. With only up to 7 sub-supernets, few-shot NAS establishes new SoTAs: on ImageNet, it finds models that reach 80.5% top-1 accuracy at 600 MB FLOPS and 77.5% top-1 accuracy at 238 MFLOPS; on CIFAR10, it reaches 98.72% top-1 accuracy without using extra data or transfer learning. In Auto-GAN, few-shot NAS outperforms the previously published results by up to 20%. Extensive experiments show that few-shot NAS significantly improves various one-shot methods, including 4 gradient-based and 6 search-based methods on 3 different tasks in NasBench-201 and NasBench1-shot-1.

Function as a Service (FaaS) has been gaining popularity as a way to deploy computations to serverless backends in the cloud. This paradigm shifts the complexity of allocating and provisioning resources to the cloud provider, which has to provide the illusion of always-available resources (i.e., fast function invocations without cold starts) at the lowest possible resource cost. Doing so requires the provider to deeply understand the characteristics of the FaaS workload. Unfortunately, there has been little to no public information on these characteristics. Thus, in this paper, we first characterize the entire production FaaS workload of Azure Functions. We show for example that most functions are invoked very infrequently, but there is an 8-order-of-magnitude range of invocation frequencies. Using observations from our characterization, we then propose a practical resource management policy that significantly reduces the number of function coldstarts,while spending fewerresources than state-of-the-practice policies., Comment: 14 pages, 20 figures. Corrected and published in USENIX ATC, July 2020. For accompanying dataset, see https://github.com/Azure/AzurePublicDataset

Neural Architecture Search (NAS) has emerged as a promising technique for automatic neural network design. However, existing MCTS based NAS approaches often utilize manually designed action space, which is not directly related to the performance metric to be optimized (e.g., accuracy), leading to sample-inefficient explorations of architectures. To improve the sample efficiency, this paper proposes Latent Action Neural Architecture Search (LaNAS), which learns actions to recursively partition the search space into good or bad regions that contain networks with similar performance metrics. During the search phase, as different action sequences lead to regions with different performance, the search efficiency can be significantly improved by biasing towards the good regions. On three NAS tasks, empirical results demonstrate that LaNAS is at least an order more sample efficient than baseline methods including evolutionary algorithms, Bayesian optimizations, and random search. When applied in practice, both one-shot and regular LaNAS consistently outperform existing results. Particularly, LaNAS achieves 99.0% accuracy on CIFAR-10 and 80.8% top1 accuracy at 600 MFLOPS on ImageNet in only 800 samples, significantly outperforming AmoebaNet with 33x fewer samples. Our code is publicly available at https://github.com/facebookresearch/LaMCTS., Comment: Accepted at TPAMI-2021

Neural Architecture Search (NAS) has shown great success in automating the design of neural networks, but the prohibitive amount of computations behind current NAS methods requires further investigations in improving the sample efficiency and the network evaluation cost to get better results in a shorter time. In this paper, we present a novel scalable Monte Carlo Tree Search (MCTS) based NAS agent, named AlphaX, to tackle these two aspects. AlphaX improves the search efficiency by adaptively balancing the exploration and exploitation at the state level, and by a Meta-Deep Neural Network (DNN) to predict network accuracies for biasing the search toward a promising region. To amortize the network evaluation cost, AlphaX accelerates MCTS rollouts with a distributed design and reduces the number of epochs in evaluating a network by transfer learning guided with the tree structure in MCTS. In 12 GPU days and 1000 samples, AlphaX found an architecture that reaches 97.84\% top-1 accuracy on CIFAR-10, and 75.5\% top-1 accuracy on ImageNet, exceeding SOTA NAS methods in both the accuracy and sampling efficiency. Particularly, we also evaluate AlphaX on NASBench-101, a large scale NAS dataset; AlphaX is 3x and 2.8x more sample efficient than Random Search and Regularized Evolution in finding the global optimum. Finally, we show the searched architecture improves a variety of vision applications from Neural Style Transfer, to Image Captioning and Object Detection., Comment: another search algorithm for NAS. arXiv admin note: substantial text overlap with arXiv:1805.07440

The performance and efficiency of distributed training of Deep Neural Networks highly depend on the performance of gradient averaging among all participating nodes, which is bounded by the communication between nodes. There are two major strategies to reduce communication overhead: one is to hide communication by overlapping it with computation, and the other is to reduce message sizes. The first solution works well for linear neural architectures, but latest networks such as ResNet and Inception offer limited opportunity for this overlapping. Therefore, researchers have paid more attention to minimizing communication. In this paper, we present a novel gradient compression framework derived from insights of real gradient distributions, and which strikes a balance between compression ratio, accuracy, and computational overhead. Our framework has two major novel components: sparsification of gradients in the frequency domain, and a range-based floating point representation to quantize and further compress gradients frequencies. Both components are dynamic, with tunable parameters that achieve different compression ratio based on the accuracy requirement and systems' platforms, and achieve very high throughput on GPUs. We prove that our techniques guarantee the convergence with a diminishing compression ratio. Our experiments show that the proposed compression framework effectively improves the scalability of most popular neural networks on a 32 GPU cluster to the baseline of no compression, without compromising the accuracy and convergence speed., Comment: 10 pages

Because robots can directly perceive and affect the physical world, security issues take on particular importance. In this paper, we describe the results of our work on scanning the entire IPv4 address space of the Internet for instances of the Robot Operating System (ROS), a widely used robotics platform for research. Our results identified that a number of hosts supporting ROS are exposed to the public Internet, thereby allowing anyone to access robotic sensors and actuators. As a proof of concept, and with consent, we were able to read image sensor information and move the robot of a research group in a US university. This paper gives an overview of our findings, including the geographic distribution of publicly-accessible platforms, the sorts of sensor and actuator data that is available, as well as the different kinds of robots and sensors that our scan uncovered. Additionally, we offer recommendations on best practices to mitigate these security issues in the future., Comment: 10 pages

Neural Architecture Search (NAS) has shown great success in automating the design of neural networks, but the prohibitive amount of computations behind current NAS methods requires further investigations in improving the sample efficiency and the network evaluation cost to get better results in a shorter time. In this paper, we present a novel scalable Monte Carlo Tree Search (MCTS) based NAS agent, named AlphaX, to tackle these two aspects. AlphaX improves the search efficiency by adaptively balancing the exploration and exploitation at the state level, and by a Meta-Deep Neural Network (DNN) to predict network accuracies for biasing the search toward a promising region. To amortize the network evaluation cost, AlphaX accelerates MCTS rollouts with a distributed design and reduces the number of epochs in evaluating a network by transfer learning, which is guided with the tree structure in MCTS. In 12 GPU days and 1000 samples, AlphaX found an architecture that reaches 97.84\% top-1 accuracy on CIFAR-10, and 75.5\% top-1 accuracy on ImageNet, exceeding SOTA NAS methods in both the accuracy and sampling efficiency. Particularly, we also evaluate AlphaX on NASBench-101, a large scale NAS dataset; AlphaX is 3x and 2.8x more sample efficient than Random Search and Regularized Evolution in finding the global optimum. Finally, we show the searched architecture improves a variety of vision applications from Neural Style Transfer, to Image Captioning and Object Detection., Comment: To appear in the Thirty-Fourth AAAI conference on Artificial Intelligence (AAAI-2020)

Index structures are one of the most important tools that DBAs leverage to improve the performance of analytics and transactional workloads. However, building several indexes over large datasets can often become prohibitive and consume valuable system resources. In fact, a recent study showed that indexes created as part of the TPC-C benchmark can account for 55% of the total memory available in a modern DBMS. This overhead consumes valuable and expensive main memory, and limits the amount of space available to store new data or process existing data. In this paper, we present FITing-Tree, a novel form of a learned index which uses piece-wise linear functions with a bounded error specified at construction time. This error knob provides a tunable parameter that allows a DBA to FIT an index to a dataset and workload by being able to balance lookup performance and space consumption. To navigate this tradeoff, we provide a cost model that helps determine an appropriate error parameter given either (1) a lookup latency requirement (e.g., 500ns) or (2) a storage budget (e.g., 100MB). Using a variety of real-world datasets, we show that our index is able to provide performance that is comparable to full index structures while reducing the storage footprint by orders of magnitude., Comment: 18 pages

Networks are difficult to configure correctly, and tricky to debug. These problems are accentuated by temporal and stateful behavior. Static verification, while useful, is ineffectual for detecting behavioral deviations induced by hardware faults, security failures, and so on, so dynamic property monitoring is also valuable. Unfortunately, existing monitoring and runtime verification for networks largely focuses on properties about individual packets (such as connectivity) or requires a digest of all network events be sent to a server, incurring enormous cost. We present a network monitoring system that avoids these problems. Because traces of network events correspond well to temporal logic, we use a subset of Metric First-Order Temporal Logic as the query language. These queries are compiled down to execute completely on the network switches. This vastly reduces network load, improves the precision of queries, and decreases detection latency. We show the practical feasibility of our work by extending a widely-used software switch and deploying it on networks. Our work also suggests improvements to network instruction sets to better support temporal monitoring.