Is the GPU Half-Empty or Half-Full? Practical Scheduling Techniques for LLMs

Serving systems for Large Language Models (LLMs) improve throughput by processing several requests concurrently. However, multiplexing hardware resources between concurrent requests involves non-trivial scheduling decisions. Practical serving systems typically implement these decisions at two levels: First, a load balancer routes requests to different servers which each hold a replica of the LLM. Then, on each server, an engine-level scheduler decides when to run a request, or when to queue or preempt it. Improved scheduling policies may benefit a wide range of LLM deployments and can often be implemented as "drop-in replacements" to a system's current policy. In this work, we survey scheduling techniques from the literature and from practical serving systems. We find that schedulers from the literature often achieve good performance but introduce significant complexity. In contrast, schedulers in practical deployments often leave easy performance gains on the table but are easy to implement, deploy and configure. This finding motivates us to introduce two new scheduling techniques, which are both easy to implement, and outperform current techniques on production workload traces.

CascadeServe: Unlocking Model Cascades for Inference Serving

Machine learning (ML) models are increasingly deployed to production, calling for efficient inference serving systems. Efficient inference serving is complicated by two challenges: (i) ML models incur high computational costs, and (ii) the request arrival rates of practical applications have frequent, high, and sudden variations which make it hard to correctly provision hardware. Model cascades are positioned to tackle both of these challenges, as they (i) save work while maintaining accuracy, and (ii) expose a high-resolution trade-off between work and accuracy, allowing for fine-grained adjustments to request arrival rates. Despite their potential, model cascades haven't been used inside an online serving system. This comes with its own set of challenges, including workload adaption, model replication onto hardware, inference scheduling, request batching, and more. In this work, we propose CascadeServe, which automates and optimizes end-to-end inference serving with cascades. CascadeServe operates in an offline and online phase. In the offline phase, the system pre-computes a gear plan that specifies how to serve inferences online. In the online phase, the gear plan allows the system to serve inferences while making near-optimal adaptations to the query load at negligible decision overheads. We find that CascadeServe saves 2-3x in cost across a wide spectrum of the latency-accuracy space when compared to state-of-the-art baselines on different workloads.

A Declarative System for Optimizing AI Workloads

Modern AI models provide the key to a long-standing dream: processing analytical queries about almost any kind of data. Until recently, it was difficult and expensive to extract facts from company documents, data from scientific papers, or insights from image and video corpora. Today's models can accomplish these tasks with high accuracy. However, a programmer who wants to answer a substantive AI-powered query must orchestrate large numbers of models, prompts, and data operations. For even a single query, the programmer has to make a vast number of decisions such as the choice of model, the right inference method, the most cost-effective inference hardware, the ideal prompt design, and so on. The optimal set of decisions can change as the query changes and as the rapidly-evolving technical landscape shifts. In this paper we present Palimpzest, a system that enables anyone to process AI-powered analytical queries simply by defining them in a declarative language. The system uses its cost optimization framework -- which explores the search space of AI models, prompting techniques, and related foundation model optimizations -- to implement the query with the best trade-offs between runtime, financial cost, and output data quality. We describe the workload of AI-powered analytics tasks, the optimization methods that Palimpzest uses, and the prototype system itself. We evaluate Palimpzest on tasks in Legal Discovery, Real Estate Search, and Medical Schema Matching. We show that even our simple prototype offers a range of appealing plans, including one that is 3.3x faster, 2.9x cheaper, and offers better data quality than the baseline method. With parallelism enabled, Palimpzest can produce plans with up to a 90.3x speedup at 9.1x lower cost relative to a single-threaded GPT-4 baseline, while obtaining an F1-score within 83.5% of the baseline. These require no additional work by the user.

Extract-Transform-Load for Video Streams

Social media, self-driving cars, and traffic cameras produce video streams at large scales and cheap cost. However, storing and querying video at such scales is prohibitively expensive. We propose to treat large-scale video analytics as a data warehousing problem: Video is a format that is easy to produce but needs to be transformed into an application-specific format that is easy to query. Analogously, we define the problem of Video Extract-Transform-Load (V-ETL). V-ETL systems need to reduce the cost of running a user-defined V-ETL job while also giving throughput guarantees to keep up with the rate at which data is produced. We find that no current system sufficiently fulfills both needs and therefore propose Skyscraper, a system tailored to V-ETL. Skyscraper can execute arbitrary video ingestion pipelines and adaptively tunes them to reduce cost at minimal or no quality degradation, e.g., by adjusting sampling rates and resolutions to the ingested content. Skyscraper can hereby be provisioned with cheap on-premises compute and uses a combination of buffering and cloud bursting to deal with peaks in workload caused by expensive processing configurations. In our experiments, we find that Skyscraper significantly reduces the cost of V-ETL ingestion compared to adaptions of current SOTA systems, while at the same time giving robustness guarantees that these systems are lacking.

RITA: Group Attention is All You Need for Timeseries Analytics

Timeseries analytics is of great importance in many real-world applications. Recently, the Transformer model, popular in natural language processing, has been leveraged to learn high quality feature embeddings from timeseries, core to the performance of various timeseries analytics tasks. However, the quadratic time and space complexities limit Transformers' scalability, especially for long timeseries. To address these issues, we develop a timeseries analytics tool, RITA, which uses a novel attention mechanism, named group attention, to address this scalability issue. Group attention dynamically clusters the objects based on their similarity into a small number of groups and approximately computes the attention at the coarse group granularity. It thus significantly reduces the time and space complexity, yet provides a theoretical guarantee on the quality of the computed attention. The dynamic scheduler of RITA continuously adapts the number of groups and the batch size in the training process, ensuring group attention always uses the fewest groups needed to meet the approximation quality requirement. Extensive experiments on various timeseries datasets and analytics tasks demonstrate that RITA outperforms the state-of-the-art in accuracy and is significantly faster -- with speedups of up to 63X.

Interpretable Outlier Summarization

Outlier detection is critical in real applications to prevent financial fraud, defend network intrusions, or detecting imminent device failures. To reduce the human effort in evaluating outlier detection results and effectively turn the outliers into actionable insights, the users often expect a system to automatically produce interpretable summarizations of subgroups of outlier detection results. Unfortunately, to date no such systems exist. To fill this gap, we propose STAIR which learns a compact set of human understandable rules to summarize and explain the anomaly detection results. Rather than use the classical decision tree algorithms to produce these rules, STAIR proposes a new optimization objective to produce a small number of rules with least complexity, hence strong interpretability, to accurately summarize the detection results. The learning algorithm of STAIR produces a rule set by iteratively splitting the large rules and is optimal in maximizing this objective in each iteration. Moreover, to effectively handle high dimensional, highly complex data sets which are hard to summarize with simple rules, we propose a localized STAIR approach, called L-STAIR. Taking data locality into consideration, it simultaneously partitions data and learns a set of localized rules for each partition. Our experimental study on many outlier benchmark datasets shows that STAIR significantly reduces the complexity of the rules required to summarize the outlier detection results, thus more amenable for humans to understand and evaluate, compared to the decision tree methods.

FactorJoin: A New Cardinality Estimation Framework for Join Queries

Cardinality estimation is one of the most fundamental and challenging problems in query optimization. Neither classical nor learning-based methods yield satisfactory performance when estimating the cardinality of the join queries. They either rely on simplified assumptions leading to ineffective cardinality estimates or build large models to understand the data distributions, leading to long planning times and a lack of generalizability across queries. In this paper, we propose a new framework FactorJoin for estimating join queries. FactorJoin combines the idea behind the classical join-histogram method to efficiently handle joins with the learning-based methods to accurately capture attribute correlation. Specifically, FactorJoin scans every table in a DB and builds single-table conditional distributions during an offline preparation phase. When a join query comes, FactorJoin translates it into a factor graph model over the learned distributions to effectively and efficiently estimate its cardinality. Unlike existing learning-based methods, FactorJoin does not need to de-normalize joins upfront or require executed query workloads to train the model. Since it only relies on single-table statistics, FactorJoin has small space overhead and is extremely easy to train and maintain. In our evaluation, FactorJoin can produce more effective estimates than the previous state-of-the-art learning-based methods, with 40x less estimation latency, 100x smaller model size, and 100x faster training speed at comparable or better accuracy. In addition, FactorJoin can estimate 10,000 sub-plan queries within one second to optimize the query plan, which is very close to the traditional cardinality estimators in commercial DBMS.

Sat2Graph: Road Graph Extraction through Graph-Tensor Encoding

Inferring road graphs from satellite imagery is a challenging computer vision task. Prior solutions fall into two categories: (1) pixel-wise segmentation-based approaches, which predict whether each pixel is on a road, and (2) graph-based approaches, which predict the road graph iteratively. We find that these two approaches have complementary strengths while suffering from their own inherent limitations. In this paper, we propose a new method, Sat2Graph, which combines the advantages of the two prior categories into a unified framework. The key idea in Sat2Graph is a novel encoding scheme, graph-tensor encoding (GTE), which encodes the road graph into a tensor representation. GTE makes it possible to train a simple, non-recurrent, supervised model to predict a rich set of features that capture the graph structure directly from an image. We evaluate Sat2Graph using two large datasets. We find that Sat2Graph surpasses prior methods on two widely used metrics, TOPO and APLS. Furthermore, whereas prior work only infers planar road graphs, our approach is capable of inferring stacked roads (e.g., overpasses), and does so robustly.

Unnatural Language Processing: Bridging the Gap Between Synthetic and Natural Language Data

Large, human-annotated datasets are central to the development of natural language processing models. Collecting these datasets can be the most challenging part of the development process. We address this problem by introducing a general purpose technique for ``simulation-to-real'' transfer in language understanding problems with a delimited set of target behaviors, making it possible to develop models that can interpret natural utterances without natural training data. We begin with a synthetic data generation procedure, and train a model that can accurately interpret utterances produced by the data generator. To generalize to natural utterances, we automatically find projections of natural language utterances onto the support of the synthetic language, using learned sentence embeddings to define a distance metric. With only synthetic training data, our approach matches or outperforms state-of-the-art models trained on natural language data in several domains. These results suggest that simulation-to-real transfer is a practical framework for developing NLP applications, and that improved models for transfer might provide wide-ranging improvements in downstream tasks.

RoadTagger: Robust Road Attribute Inference with Graph Neural Networks

Inferring road attributes such as lane count and road type from satellite imagery is challenging. Often, due to the occlusion in satellite imagery and the spatial correlation of road attributes, a road attribute at one position on a road may only be apparent when considering far-away segments of the road. Thus, to robustly infer road attributes, the model must integrate scattered information and capture the spatial correlation of features along roads. Existing solutions that rely on image classifiers fail to capture this correlation, resulting in poor accuracy. We find this failure is caused by a fundamental limitation -- the limited effective receptive field of image classifiers. To overcome this limitation, we propose RoadTagger, an end-to-end architecture which combines both Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs) to infer road attributes. The usage of graph neural networks allows information propagation on the road network graph and eliminates the receptive field limitation of image classifiers. We evaluate RoadTagger on both a large real-world dataset covering 688 km^2 area in 20 U.S. cities and a synthesized micro-dataset. In the evaluation, RoadTagger improves inference accuracy over the CNN image classifier based approaches. RoadTagger also demonstrates strong robustness against different disruptions in the satellite imagery and the ability to learn complicated inductive rules for aggregating scattered information along the road network.