Mixture of Experts (MoE) LLMs, with sparse activation patterns, offer a promising approach to scaling language models while avoiding proportional cost increases. However, their large parameter sizes pose deployment challenges in resource-constrained environments with limited GPU memory. Typical deployments use CPU-GPU hybrid execution, where the GPU handles compute-intensive GEMM operations and the CPU processes the attention mechanism. This setup introduces a challenge: optimizing resource utilization across CPU and GPU. Prior work designs system optimizations based on performance models with a limited scope that don’t capture complex hardware-system interactions. Therefore, they neither identify nor achieve hardware limits.

Workload execution can account for 90% of the total database knob tuning time, which is often the bottleneck for efficient knob tuning in practice. Reducing the tuning time by using a compressed workload is a natural solution. However, many existing workload compression methods are designed for OLAP workloads, which reduce the number of queries needed for analysis tasks by sampling a small subset of queries. These methods are less effective for OLTP workloads in knob-tuning tasks, as they often disregard essential contextual details, including query sequence and concurrency. As a result, configurations that perform well on the compressed OLTP workload may not deliver similar competitive performance on the original workload. To address these challenges, we first define the objective of OLTP workload compression for knob tuning. We then propose a slice-based compression method, SCompression , which compresses workloads by slicing based on time intervals while preserving concurrency. SCompression achieves the objective by focusing on generating a compressed workload that (1) executes faster than the original workload and (2) produces performance variations similar to the source workload under different configurations. SCompression works in three steps: (1) dividing the workload into segments to capture regular performance fluctuations, (2) slicing each segment to preserve concurrency and transaction context, and (3) sampling slices under execution time constraints using a cluster-based approach to ensure representativeness. Finally, SCompression replays the compressed workload to produce the performance that mirrors the source workload. Extensive experiments on real-world and benchmark OLTP workloads show that SCompression is a cost-effective solution for knob tuning, accelerating tuning by up to 40× with only a 5% performance reduction.

Autonomous database management systems (DBMSs) aim to optimize themselves automatically without human guidance. They rely on machine learning (ML) models that predict their run-time behavior to evaluate whether a candidate configuration is beneficial without the expensive execution of queries. However, the high cost of collecting the training data to build these models makes them impractical for real-world deployments. Furthermore, these models are instance-specific and thus require retraining whenever the DBMS's environment changes. State-of-the-art methods spend over 93% of their time running queries for training versus tuning. To mitigate this problem, we present the Boot framework for automatically accelerating training data collection in DBMSs. Boot utilizes macro- and micro-acceleration (MMA) techniques that modify query execution semantics with approximate run-time telemetry and skip repetitive parts of the training process. To evaluate Boot, we integrated it into a database gym for PostgreSQL. Our experimental evaluation shows that Boot reduces training collection times by up to 268× with modest degradation in model accuracy. These results also indicate that our MMA-based approach scales with dataset size and workload complexity.

Despite the high computational throughput of GPUs, limited memory capacity and bandwidth-limited CPU-GPU communication via PCIe links remain significant bottlenecks for accelerating large-scale data analytics workloads. This paper introduces Vortex, a GPU-accelerated framework designed for data analytics workloads that exceed GPU memory capacity. A key aspect of our framework is an optimized IO primitive that leverages all available PCIe links in multi-GPU systems for the IO demand of a single target GPU. It routes data through other GPUs to such target GPU that handles IO-intensive analytics tasks. This approach is advantageous when other GPUs are occupied with compute-bound workloads, such as popular AI applications that typically underutilize IO resources. We also introduce a novel programming model that separates GPU kernel development from IO scheduling, reducing programmer burden and enabling GPU code reuse. Additionally, we present the design of certain important query operators and discuss a late materialization technique based on GPU's zero-copy memory access. Without caching any data in GPU memory, Vortex improves the performance of the state-of-the-art GPU baseline, Proteus, by 5.7× on average and enhances price performance by 2.5× compared to a CPU-based DuckDB baseline.

A self-driving database management system (DBMS) aims to configure, deploy, and optimize almost all aspects of itself automatically without human intervention or guidance. Achieving this high level of automation relies on machine learning (ML) models that predict how a DBMS will behave in different scenarios. This behavior encompasses all DBMS runtime operations, including query execution and maintenance tasks. These ML-based behavior models for a self-driving DBMS require low-level training data about a DBMS's internals. Such training data includes (1) features that describe the workload, environment, and DBMS configuration, and (2) both DBMS- and hardware-level metrics. But it is difficult to collect training data from a DBMS while it is running because it can introduce performance and measurement degradations that hinder the ML models' ability to predict the DBMS's behavior correctly. We present the TScout (TS) framework for collecting training data from self-driving DBMSs. Our framework is an internal approach where developers annotate a DBMS's source code with hooks to monitor the system's behavior. TS then extracts these hooks and generates a kernel-level program (via Linux's BPF) that efficiently captures metrics from multiple sources (e.g., CPU performance counters, memory allocators). TS combines these metrics with internal DBMS state observations, generating training data for behavior models. We integrated TS in a PostgreSQL-compatible DBMS and measured its ability to collect training data for both OLTP and OLAP workloads. Our results show that TS generates training data for a deployed DBMS to train more accurate models than previous methods with only a 7% performance reduction.

Advances in memory technology have made it feasible for database management systems (DBMS) to store their working data set in main memory. This trend shifts the bottleneck for query execution from disk accesses to CPU efficiency. One technique to improve CPU efficiency is batch-oriented processing, or vectorization, as it reduces interpretation overhead. For each vector (batch) of tuples, the DBMS must track the set of valid (visible) tuples that survive all previous processing steps. To that end, existing systems employ one of two data structures, or filter representations: selection vectors or bitmaps. In this work, we analyze each approach's strengths and weaknesses and offer recommendations on how to implement vectorized operations. Through a wide range of micro-benchmarks, we determine that the optimal strategy is a function of many factors: the cost of iterating through tuples, the cost of the operation itself, and how amenable it is to SIMD vectorization. Our analysis shows that bitmaps perform better for operations that can be vectorized using SIMD instructions and that selection vectors perform better on all other operations due to cheaper iteration logic.

Database management systems (DBMSs) are notoriously difficult to deploy and administer. The goal of a self-driving DBMS is to remove these impediments by managing itself automatically. However, a critical problem in achieving full autonomy is how to predict the DBMS's runtime behavior and resource consumption. These predictions guide a self-driving DBMS's decision-making components to tune and optimize all aspects of the system. We present the ModelBot2 end-to-end framework for constructing and maintaining prediction models using machine learning (ML) in self-driving DBMSs. Our approach decomposes a DBMS's architecture into fine-grained operating units that make it easier to estimate the system's behavior for configurations that it has never seen before. ModelBot2 then provides an offline execution environment to exercise the system to produce the training data used to train its models. We integrated ModelBot2 in an in-memory DBMS and measured its ability to predict its performance for OLTP and OLAP workloads running in dynamic environments. We also compare ModelBot2 against state-of-the-art ML models and show that our models are up to 25x more accurate in multiple scenarios.

Database management systems (DBMSs) are notoriously difficult to deploy and administer. Self-driving DBMSs seek to remove these impediments by managing themselves automatically. Despite decades of DBMS auto-tuning research, a truly autonomous, self-driving DBMS is yet to come. But recent advancements in artificial intelligence and machine learning (ML) have moved this goal closer. Given this, we present a system implementation treatise towards achieving a self-driving DBMS. We first provide an overview of the NoisePage self-driving DBMS that uses ML to predict the DBMS's behavior and optimize itself without human support or guidance. The system's architecture has three main ML-based components: (1) workload forecasting, (2) behavior modeling, and (3) action planning. We then describe the system design principles to facilitate holistic autonomous operations. Such prescripts reduce the complexity of the problem, thereby enabling a DBMS to converge to a better and more stable configuration more quickly.

Recent research has shown promising results by using machine learning (ML) techniques to improve the performance of database systems, e.g., in query optimization or index recommendation. However, in many production deployments, the ML models' performance degrades significantly when the test data diverges from the data used to train these models. In this paper, we address this performance degradation by using B-instances to collect additional data during deployment. We propose an active data collection platform, ADCP, that employs active learning (AL) to gather relevant data cost-effectively. We develop a novel AL technique, Holistic Active Learner (HAL), that robustly combines multiple noisy signals for data gathering in the context of database applications. HAL applies to various ML tasks, budget sizes, cost types, and budgeting interfaces for database applications. We evaluate ADCP on both industry-standard benchmarks and real customer workloads. Our evaluation shows that, compared with other baselines, our technique improves ML models' prediction performance by up to 2x with the same cost budget. In particular, on production workloads, our technique reduces the prediction error of ML models by 75% using about 100 additionally collected queries.

Just-in-time (JIT) query compilation is a technique to improve analytical query performance in database management systems (DBMSs). But the cost of compiling each query can be significant relative to its execution time. This overhead prohibits the DBMS from employing well-known adaptive query processing (AQP) methods to generate a new plan for a query if data distributions do not match the optimizer's estimations. The optimizer could eagerly generate multiple sub-plans for a query, but it can only include a few alternatives as each addition increases the compilation time. We present a method, called Permutable Compiled Queries (PCQ), that bridges the gap between JIT compilation and AQP. It allows the DBMS to modify compiled queries without needing to recompile or including all possible variations before the query starts. With PCQ, the DBMS structures a query's code with indirection layers that enable the DBMS to change the plan even while it is running. We implement PCQ in an in-memory DBMS and compare it against non-adaptive plans in a microbenchmark and against state-of-the-art analytic DBMSs. Our evaluation shows that PCQ outperforms static plans by more than 4X and yields better performance on an analytical benchmark by more than 2X against other DBMSs.

The first step towards an autonomous database management system (DBMS) is the ability to model the target application's workload. This is necessary to allow the system to anticipate future workload needs and select the proper optimizations in a timely manner. Previous forecasting techniques model the resource utilization of the queries. Such metrics, however, change whenever the physical design of the database and the hardware resources change, thereby rendering previous forecasting models useless.

In-memory database management systems (DBMSs) outperform disk-oriented systems for on-line transaction processing (OLTP) workloads. But this improved performance is only achievable when the database is smaller than the amount of physical memory available in the system. To overcome this limitation, some in-memory DBMSs can move cold data out of volatile DRAM to secondary storage. Such data appears as if it resides in memory with the rest of the database even though it does not.

Using indexes for query execution is crucial for achieving high performance in modern on-line transaction processing databases. For a main-memory database, however, these indexes consume a large fraction of the total memory available and are thus a major source of storage overhead of in-memory databases. To reduce this overhead, we propose using a two-stage index: The first stage ingests all incoming entries and is kept small for fast read and write operations. The index periodically migrates entries from the first stage to the second, which uses a more compact, read-optimized data structure. Our first contribution is hybrid index, a dual-stage index architecture that achieves both space efficiency and high performance. Our second contribution is Dual-Stage Transformation (DST), a set of guidelines for converting any order-preserving index structure into a hybrid index. Our third contribution is applying DST to four popular order-preserving index structures and evaluating them in both standalone microbenchmarks and a full in-memory DBMS using several transaction processing workloads. Our results show that hybrid indexes provide comparable throughput to the original ones while reducing the memory overhead by up to 70%.

Graph partition quality affects the overall performance of parallel graph computation systems. The quality of a graph partition is measured by the balance factor and edge cut ratio. A balanced graph partition with small edge cut ratio is generally preferred since it reduces the expensive network communication cost. However, according to an empirical study on Giraph, the performance over well partitioned graph might be even two times worse than simple random partitions. This is because these systems only optimize for the simple partition strategies and cannot efficiently handle the increasing workload of local message processing when a high quality graph partition is used. In this paper, we propose a novel partition aware graph computation engine named PAGE, which equips a new message processor and a dynamic concurrency control model. The new message processor concurrently processes local and remote messages in a unified way. The dynamic model adaptively adjusts the concurrency of the processor based on the online statistics. The experimental evaluation demonstrates the superiority of PAGE over the graph partitions with various qualities.

Subgraph listing is a fundamental operation to many graph and network analyses. The problem itself is computationally expensive and is well-studied in centralized processing algorithms. However, the centralized solutions cannot scale well to large graphs. Recently, several parallel approaches are introduced to handle the large graphs. Unfortunately, these parallel approaches still rely on the expensive join operations, thus cannot achieve high performance. In this paper, we design a novel parallel subgraph listing framework, named PSgL. The PSgL iteratively enumerates subgraph instances and solves the subgraph listing in a divide-and-conquer fashion. The framework completely relies on the graph traversal, and avoids the explicit join operation. Moreover, in order to improve its performance, we propose several solutions to balance the workload and reduce the size of intermediate results. Specially, we prove the problem of partial subgraph instance distribution for workload balance is NP-hard, and carefully design a set of heuristic strategies. To further reduce the enormous intermediate results, we introduce three independent mechanisms, which are automorphism breaking of the pattern graph, initial pattern vertex selection based on a cost model, and a pruning method based on a light-weight index.

Graph partitioning is one of the key components in parallel graph computation, and the partition quality significantly affects the overall computing performance. In the existing graph computing systems, ``good'' partition schemes are preferred as they have smaller edge cut ratio and hence reduce the communication cost among working nodes. However, in an empirical study on Giraph[1], we found that the performance over well partitioned graph might be even two times worse than simple partitions. The cause is that the local message processing cost in graph computing systems may surpass the communication cost in several cases.