Small deep learning models deployed at edge servers suffer from decreasing accuracy due to data drift and hence require continual learning, which leads to resource competition with inference execution, decreasing accuracy and latency service-level-objective (SLO) fulfillment. Previous methods fail to maximize both accuracy and SLO fulfillment. To address this problem, we propose a Checkpointed Inference-based system for accurate and SLO-guaranteed data drift-resilient model Serving (CIS). CIS incorporates checkpointed inference - maximizing accuracy by continuously switching to intermediately retrained models during inference. However, model switching introduces significant inference latency overhead. To mitigate this, first, CIS proposes a lightweight allocation and placement scheduler to minimize switching time. Second, during inference execution, CIS reorders retraining samples to reduce switching frequency. Third, it temporarily reallocates GPU space from retraining tasks to inference tasks to address request queuing issue caused by the switching, with minimal impact on accuracy. Trace-driven experiments show that CIS achieves up to 25.1% higher accuracy without affecting SLO fulfillment, and requires 4× lower GPU cost compared to the existing method in achieving similar or higher accuracy.

RAG (Retrieval Augmented Generation) allows LLMs (large language models) to generate better responses with external knowledge, but using more external knowledge causes higher response delay. Prior work focuses either on reducing the response delay (e.g., better scheduling of RAG queries) or on maximizing quality (e.g., tuning the RAG workflow), but they fall short in systematically balancing the tradeoff between the delay and quality of RAG responses. To balance both quality and response delay, this paper presents METIS, the first RAG system that jointly schedules queries and adapts the key RAG configurations of each query, such as the number of retrieved text chunks and synthesis methods. Using four popular RAG-QA datasets, we show that compared to the state-of-the-art RAG optimization schemes, METIS reduces the generation latency by 1.64 – 2.54× without sacrificing generation quality.

As large language models (LLMs) take on complex tasks, their inputs are supplemented with longer contexts that incorporate domain knowledge. Yet using long contexts is challenging as nothing can be generated until the whole context is processed by the LLM. While the context-processing delay can be reduced by reusing the KV cache of a context across different inputs, fetching the KV cache, which contains large tensors, over the network can cause high extra network delays.

Deep learning inference on streaming media data, such as object detection in video or LiDAR feeds and text extraction from audio waves, is now ubiquitous. To achieve high inference accuracy, these applications typically require significant network bandwidth to gather high-fidelity data and extensive GPU resources to run deep neural networks (DNNs). While the high demand for network bandwidth and GPU resources could be substantially reduced by optimally adapting the configuration knobs, such as video resolution and frame rate, current adaptation techniques fail to meet three requirements simultaneously: adapt configurations (i) with minimum extra GPU or bandwidth overhead (ii) to reach near-optimal decisions based on how the data affects the final DNN's accuracy, and (iii) do so for a range of configuration knobs. This paper presents OneAdapt, which meets these requirements by leveraging a gradient-ascent strategy to adapt configuration knobs. The key idea is to embrace DNNs' differentiability to quickly estimate the accuracy's gradient to each configuration knob, called AccGrad. Specifically, OneAdapt estimates AccGrad by multiplying two gradients: InputGrad (i.e., how each configuration knob affects the input to the DNN) and DNNGrad (i.e., how the DNN input affects the DNN inference output). We evaluate OneAdapt across five types of configurations, four analytic tasks, and five types of input data. Compared to state-of-the-art adaptation schemes, OneAdapt cuts bandwidth usage and GPU usage by 15-59% while maintaining comparable accuracy or improves accuracy by 1-5% while using equal or fewer resources.

As Video Streaming and Analytics (VSA) systems become increasingly popular, serious privacy concerns have risen on exposing too much unnecessary private information to the VSA providers. Yet, it is challenging to protect privacy while still preserving desired VSA features, i.e., effective analytics, forensic support, resource efficiency, and real-time execution. In this paper, we present a VSA privacy enhancement system (PECAM), which addresses the above challenge with no change in the VSA back-end. PECAM leverages a novel Generative Adversarial Network to perform the privacy-enhanced securely-reversible video transformation. PECAM also incorporates a couple of system optimizations into its VSA workflow to reduce network bandwidth usage and enable real-time processing on cameras. We implement our PECAM prototype on commodity hardware and evaluate its performance via both security study and extensive experiments. Results demonstrate that PECAM can effectively enhance the visual privacy of VSA in the presence of an adversary, and its transformed videos, when taken as input for various VSA back-end tasks, maintain a 96% accuracy of corresponding original videos. Additionally, it performs 12.3× and 1.8× better than baseline methods in terms of the computing cost and network bandwidth usage, respectively.

The conventional wisdom requires that all congestion control algorithms deployed on the public Internet be TCP-friendly. If universally obeyed, this requirement would greatly constrain the future of such congestion control algorithms. If partially ignored, as is increasingly likely, then there could be significant inequities in the bandwidth received by different flows. To avoid this dilemma, we propose an alternative to the TCP-friendly paradigm that can accommodate innovation, is consistent with the Internet's current economic model, and is feasible to deploy given current usage trends.

MLaaS (ML-as-a-Service) offerings by cloud computing platforms are becoming increasingly popular. Hosting pre-trained machine learning models in the cloud enables elastic scalability as the demand grows. But providing low latency and reducing the latency variance is a key requirement. Variance is harder to control in a cloud deployment due to uncertain-ties in resource allocations across many virtual instances. We propose the collage inference technique, which uses a novel convolutional neural network model, collage-cnn, to provide low-cost redundancy. A collage-cnn model takes a collage image formed by combining multiple images and performs multi-image classification in one shot, albeit at slightly lower accuracy. We augment a collection of traditional single image classifier models with a single collage-cnn classifier, which acts as their low-cost redundant backup. Collage-cnn provides backup classification results if any single image classification requests experience a slowdown. Deploying the collage-cnn models in the cloud, we demonstrate that the 99th percentile tail latency of inference can be reduced by 1.2x to 2x compared to replication-based approaches while providing high accuracy. Variation in inference latency can be reduced by 1.8x to 15x.

Cameras are deployed at scale with the purpose of searching and tracking objects of interest (e.g., a suspected person) through the camera network on live videos. Such cross-camera analytics is data and compute intensive, whose costs grow with the number of cameras and time. We present Spatula, a cost-efficient system that enables scaling cross-camera analytics on edge compute boxes to large camera networks by leveraging the spatial and temporal cross-camera correlations. While such correlations have been used in computer vision community, Spatula uses them to drastically reduce the communication and computation costs by pruning search space of a query identity (e.g., ignoring frames not correlated with the query identity’s current position). Spatula provides the first system substrate on which cross-camera analytics applications can be built to efficiently harness the cross-camera correlations that are abundant in large camera deployments. Spatula reduces compute load by $8.3\times$ on an 8-camera dataset, and by $23\times-86\times$ on two datasets with hundreds of cameras (simulated from real vehicle/pedestrian traces). We have also implemented Spatula on a testbed of 5 AWS DeepLens cameras.

No abstract available.

The world is witnessing an unprecedented increase in camera deployment. The USA and UK, for instance, have one camera for every 8 people. Video analytics from these cameras are becoming more and more pervasive, exerting important functions on a wide range of verticals including manufacturing, transportation, and retails. While vision techniques have seen considerable advancement, they have come at the expense of compute and network cost.

The network communications between the cloud and the client have become the weak link for global cloud services that aim to provide low latency services to their clients. In this paper, we first characterize WAN latency from the viewpoint of a large cloud provider Azure, whose network edges serve hundreds of billions of TCP connections a day across hundreds of locations worldwide. In particular, we focus on instances of latency degradation and design a tool, BlameIt, that enables cloud operators to localize the cause (i.e., faulty AS) of such degradation. BlameIt uses passive diagnosis, using measurements of existing connections between clients and the cloud locations, to localize the cause to one of cloud, middle, or client segments. Then it invokes selective active probing (within a probing budget) to localize the cause more precisely. We validate BlameIt by comparing its automatic fault localization results with that arrived at by network engineers manually, and observe that BlameIt correctly localized the problem in all the 88 incidents. Further, BlameIt issues 72X fewer active probes than a solution relying on active probing alone, and is deployed in production at Azure.

Running data-parallel jobs across geo-distributed sites has emerged as a promising direction due to the growing need for geo-distributed cluster deployment. A key difference between geo-distributed and intra-cluster jobs is the heterogeneous (and often constrained) nature of compute and network resources across the sites. We propose Tetrium, a system for multi-resource allocation in geo-distributed clusters, that jointly considers both compute and network resources for task placement and job scheduling. Tetrium significantly reduces job response time, while incorporating several other performance goals with simple control knobs. Our EC2 deployment and trace-driven simulations suggest that Tetrium improves the average job response time by up to 78% compared to existing data-locality-based solutions, and up to 55% compared to Iridium, the recently proposed geo-distributed analytics system.

Applying deep convolutional neural networks (NN) to video data at scale poses a substantial systems challenge, as improving inference accuracy often requires a prohibitive cost in computational resources. While it is promising to balance resource and accuracy by selecting a suitable NN configuration (e.g., the resolution and frame rate of the input video), one must also address the significant dynamics of the NN configuration's impact on video analytics accuracy. We present Chameleon, a controller that dynamically picks the best configurations for existing NN-based video analytics pipelines. The key challenge in Chameleon is that in theory, adapting configurations frequently can reduce resource consumption with little degradation in accuracy, but searching a large space of configurations periodically incurs an overwhelming resource overhead that negates the gains of adaptation. The insight behind Chameleon is that the underlying characteristics (e.g., the velocity and sizes of objects) that affect the best configuration have enough temporal and spatial correlation to allow the search cost to be amortized over time and across multiple video feeds. For example, using the video feeds of five traffic cameras, we demonstrate that compared to a baseline that picks a single optimal configuration offline, Chameleon can achieve 20-50% higher accuracy with the same amount of resources, or achieve the same accuracy with only 30--50% of the resources (a 2-3X speedup).

Systems are increasingly required to provide responses to queries, even if not exact, within stringent time deadlines. These systems parallelize computations over many processes and aggregate them hierarchically to get the final response (e.g., search engines and data analytics). Due to large performance variations in clusters, some processes are slower. Therefore, aggregators are faced with the question of how long to wait for outputs from processes before combining and sending them upstream. Longer waits increase the response quality as it would include outputs from more processes. However, it also increases the risk of the aggregator failing to provide its result by the deadline. This leads to all its results being ignored, degrading response quality. Our algorithm, Cedar, proposes a solution to this quandary of deciding wait durations at aggregators. It uses an online algorithm to learn distributions of durations at each level in the hierarchy and collectively optimizes the wait duration. Cedar's solution is theoretically sound, fully distributed, and generically applicable across systems that use aggregation trees since it is agnostic to the causes of performance variations. Evaluation using production latency distributions from Google, Microsoft and Facebook using deployment and simulation shows that Cedar improves average response quality by over 100%.

Interactive real-time streaming applications such as audio-video conferencing, online gaming and app streaming, place stringent requirements on the network in terms of delay, jitter, and packet loss. Many of these applications inherently involve client-to-client communication, which is particularly challenging since the performance requirements need to be met while traversing the public wide-area network (WAN). This is different from the typical situation of cloud-to-client communication, where the WAN can often be bypassed by moving a communication end-point to a cloud “edge”, close to the client. Can we nevertheless take advantage of cloud resources to improve the performance of real-time client-to-client streaming over the WAN?

The drive towards richer and more interactive web content places increasingly stringent requirements on datacenter network performance. Applications running atop these networks typically partition an incoming query into multiple subqueries, and generate the final result by aggregating the responses for these subqueries. As a result, a large fraction --- as high as 80% --- of the network flows in such workloads are short and latency-sensitive. The speed with which existing networks respond to packet drops limits their ability to meet high-percentile flow completion time SLOs. Indirect notifications indicating packet drops (e.g., duplicates in an end-to-end acknowledgement sequence) are an important limitation to the agility of response to packet drops.

Low latency analytics on geographically distributed datasets (across datacenters, edge clusters) is an upcoming and increasingly important challenge. The dominant approach of aggregating all the data to a single datacenter significantly inflates the timeliness of analytics. At the same time, running queries over geo-distributed inputs using the current intra-DC analytics frameworks also leads to high query response times because these frameworks cannot cope with the relatively low and variable capacity of WAN links. We present Iridium, a system for low latency geo-distributed analytics. Iridium achieves low query response times by optimizing placement of both data and tasks of the queries. The joint data and task placement optimization, however, is intractable. Therefore, Iridium uses an online heuristic to redistribute datasets among the sites prior to queries' arrivals, and places the tasks to reduce network bottlenecks during the query's execution. Finally, it also contains a knob to budget WAN usage. Evaluation across eight worldwide EC2 regions using production queries show that Iridium speeds up queries by 3× -- 19× and lowers WAN usage by 15% -- 64% compared to existing baselines.

As clusters continue to grow in size and complexity, providing scalable and predictable performance is an increasingly important challenge. A crucial roadblock to achieving predictable performance is stragglers, i.e., tasks that take significantly longer than expected to run. At this point, speculative execution has been widely adopted to mitigate the impact of stragglers. However, speculation mechanisms are designed and operated independently of job scheduling when, in fact, scheduling a speculative copy of a task has a direct impact on the resources available for other jobs. In this work, we present Hopper, a job scheduler that is speculation-aware, i.e., that integrates the tradeoffs associated with speculation into job scheduling decisions. We implement both centralized and decentralized prototypes of the Hopper scheduler and show that 50% (66%) improvements over state-of-the-art centralized (decentralized) schedulers and speculation strategies can be achieved through the coordination of scheduling and speculation.

Straggler tasks continue to be a major hurdle in achieving faster completion of data intensive applications running on modern data-processing frameworks. Existing straggler mitigation techniques are inefficient due to their reactive and replicative nature -- they rely on a wait-speculate-re-execute mechanism, thus leading to delayed straggler detection and inefficient resource utilization. Existing proactive techniques also over-utilize resources due to replication. Existing modeling-based approaches are hard to rely on for production-level adoption due to modeling errors. We present Wrangler, a system that proactively avoids situations that cause stragglers. Wrangler automatically learns to predict such situations using a statistical learning technique based on cluster resource utilization counters. Furthermore, Wrangler introduces a notion of a confidence measure with these predictions to overcome the modeling error problems; this confidence measure is then exploited to achieve a reliable task scheduling. In particular, by using these predictions to balance delay in task scheduling against the potential for idling of resources, Wrangler achieves a speed up in the overall job completion time. For production-level workloads from Facebook and Cloudera's customers, Wrangler improves the 99th percentile job completion time by up to 61% as compared to speculative execution, a widely used straggler mitigation technique. Moreover, Wrangler achieves this speed-up while significantly improving the resource consumption (by up to 55%).

Tasks in modern data parallel clusters have highly diverse resource requirements, along CPU, memory, disk and network. Any of these resources may become bottlenecks and hence, the likelihood of wasting resources due to fragmentation is now larger. Today's schedulers do not explicitly reduce fragmentation. Worse, since they only allocate cores and memory, the resources that they ignore (disk and network) can be over-allocated leading to interference, failures and hogging of cores or memory that could have been used by other tasks. We present Tetris, a cluster scheduler that packs, i.e., matches multi-resource task requirements with resource availabilities of machines so as to increase cluster efficiency (makespan). Further, Tetris uses an analog of shortest-running-time-first to trade-off cluster efficiency for speeding up individual jobs. Tetris' packing heuristics seamlessly work alongside a large class of fairness policies. Trace-driven simulations and deployment of our prototype on a 250 node cluster shows median gains of 30% in job completion time while achieving nearly perfect fairness.

Data-intensive computing (DISC) frameworks scale by partitioning a job across a set of fault-tolerant tasks, then diffusing those tasks across large clusters. Multi-tenanted clusters must accommodate service-level objectives (SLO) in their resource model, often expressed as a maximum latency for allocating the desired set of resources to every job. When jobs are partitioned into tasks statically, a cluster cannot meet its SLOs while maintaining both high utilization and efficiency. Ideally, we want to give resources to jobs when they are free but would expect to reclaim them instantaneously when new jobs arrive, without losing work. DISC frameworks do not support such elasticity because interrupting running tasks incurs high overheads. Amoeba enables lightweight elasticity in DISC frameworks by identifying points at which running tasks of over-provisioned jobs can be safely exited, committing their outputs, and spawning new tasks for the remaining work. Effectively, tasks of DISC jobs are now sized dynamically in response to global resource scarcity or abundance. Simulation and deployment of our prototype shows that Amoeba speeds up jobs by 32% without compromising utilization or efficiency.

To improve data availability and resilience MapReduce frameworks use file systems that replicate data uniformly. However, analysis of job logs from a large production cluster shows wide disparity in data popularity. Machines and racks storing popular content become bottlenecks; thereby increasing the completion times of jobs accessing this data even when there are machines with spare cycles in the cluster. To address this problem, we present Scarlett, a system that replicates blocks based on their popularity. By accurately predicting file popularity and working within hard bounds on additional storage, Scarlett causes minimal interference to running jobs. Trace driven simulations and experiments in two popular MapReduce frameworks (Hadoop, Dryad) show that Scarlett effectively alleviates hotspots and can speed up jobs by 20.2%.

Mobile devices are increasingly equipped with hardware and software services allowing them to determine their locations, but support for building location-aware applications remains rudimentary. This paper proposes tracks of location coordinates as a high-level abstraction for a new class of mobile applications including ride sharing, location-based collaboration, and health monitoring. Each track is a sequence of entries recording a person's time, location, and application-specific data. StarTrack provides applications with a comprehensive set of operations for recording, comparing, clustering and querying tracks. StarTrack can efficiently operate on thousands of tracks.

Mobile devices are increasingly equipped with multiple network interfaces with complementary characteristics. In particular, the Wi-Fi interface has high throughput and transfer power efficiency, but its idle power consumption is prohibitive. In this paper we present, Blue-Fi, a sytem that predicts the availability of the Wi-Fi connectivity by using a combination of bluetooth contact-patterns and cell-tower information. This allows the device to intelligently switch the Wi-Fi interface on only when there is Wi-Fi connectivity available, thus avoiding the long periods in idle state and significantly reducing the the number of scans for discovery.

Mobile devices are increasingly equipped with multiple network interfaces: Wireless Local Area Network (WLAN) interfaces for local connectivity and Wireless Wide Area Network (WWAN) interfaces for wide-area connectivity. The WWAN typically provides much wider coverage but much lower speeds than the WLAN. To address this dichotomy, we present COMBINE, a system for collaborative downloading wherein devices that are within WLAN range pool together their WWAN links, significantly increasing the effective speed available to them.