Sebastian Breß

Why it is time for a HyPE: a hybrid query processing engine for efficient GPU coprocessing in DBMS

GPU acceleration is a promising approach to speed up query processing of database systems by using low cost graphic processors as coprocessors. Two major trends have emerged in this area: (1) The development of frameworks for scheduling tasks in heterogeneous CPU/GPU platforms, which is mainly in the context of coprocessing for applications and does not consider specifics of database-query processing and optimization. (2) The acceleration of database operations using efficient GPU algorithms, which typically cannot be applied easily on other database systems, because of their analytical-algorithm-specific cost models. One major challenge is how to combine traditional database query processing with GPU coprocessing techniques and efficient database operation scheduling in a GPU-aware query optimizer. In this thesis, we develop a hybrid query processing engine, which extends the traditional physical optimization process to generate hybrid query plans and to perform a cost-based optimization in a way that the advantages of CPUs and GPUs are combined. Furthermore, we aim at a portable solution between different GPU-accelerated database management systems to maximize applicability. Preliminary results indicate great potential.

The Design and Implementation of CoGaDB: A Column-oriented GPU-accelerated DBMS

Nowadays, the performance of processors is primarily bound by a fixed energy budget, the power wall. This forces hardware vendors to optimize processors for specific tasks, which leads to an increasingly heterogeneous hardware landscape. Although efficient algorithms for modern processors such as GPUs are heavily investigated, we also need to prepare the database optimizer to handle computations on heterogeneous processors. GPUs are an interesting base for case studies, because they already offer many difficulties we will face tomorrow.In this paper, we present CoGaDB, a main-memory DBMS with built-in GPU acceleration, which is optimized for OLAP workloads. CoGaDB uses the self-tuning optimizer framework HyPE to build a hardware-oblivious optimizer, which learns cost models for database operators and efficiently distributes a workload on available processors. Furthermore, CoGaDB implements efficient algorithms on CPU and GPU and efficiently supports star joins. We show in this paper, how these novel techniques interact with each other in a single system. Our evaluation shows that CoGaDB quickly adapts to the underlying hardware by increasing the accuracy of its cost models at runtime.

GPU-Accelerated Database Systems: Survey and Open Challenges

The vast amount of processing power and memory bandwidth provided by modern graphics cards make them an interesting platform for data-intensive applications. Unsurprisingly, the database research community identified GPUs as effective co-processors for data processing several years ago. In the past years, there were many approaches to make use of GPUs at different levels of a database system. In this paper, we explore the design space of GPU-accelerated database management systems. Based on this survey, we present key properties, important trade-offs and typical challenges of GPU-aware database architectures, and identify major open challenges. Additionally, we survey existing GPU-accelerated DBMSs and classify their architectural properties. Then, we summarize typical optimizations implemented in GPU-accelerated DBMSs. Finally, we propose a reference architecture, indicating how GPU acceleration can be integrated in existing DBMSs.

Robust Query Processing in Co-Processor-accelerated Databases

Technology limitations are making the use of heterogeneous computing devices much more than an academic curiosity. In fact, the use of such devices is widely acknowledged to be the only promising way to achieve application-speedups that users urgently need and expect. However, building a robust and efficient query engine for heterogeneous co-processor environments is still a significant challenge. In this paper, we identify two effects that limit performance in case co-processor resources become scarce. Cache thrashing occurs when the working set of queries does not fit into the co-processors data cache, resulting in performance degradations up to a factor of 24. Heap contention occurs when multiple operators run in parallel on a co-processor and when their accumulated memory footprint exceeds the main memory capacity of the co-processor, slowing down query execution by up to a factor of six. We propose solutions for both effects. Data-driven operator placement avoids data movements when they might be harmful; query chopping limits co-processor memory usage and thus avoids contention. The combined approach-data-driven query chopping-achieves robust and scalable performance on co-processors. We validate our proposal with our open-source GPU-accelerated database engine CoGaDB and the popular star schema and TPC-H benchmarks.

Automatic selection of processing units for coprocessing in databases

Specialized processing units such as GPUs or FPGAs provide great opportunities to speed up database operations by exploiting parallelism and relieving the CPU. But utilizing coprocessors efficiently poses major challenges to developers. Besides finding fine-granular data parallel algorithms and tuning them for the available hardware, it has to be decided at runtime which (co)processor should be chosen to execute a specific task. Depending on input parameters, wrong decisions can lead to severe performance degradations since involving coprocessors introduces a significant overhead, e.g., for data transfers. In this paper, we present a framework that automatically learns and adapts execution models for arbitrary algorithms on any (co)processor to find break-even points and support scheduling decisions. We demonstrate its applicability for three common use cases in modern database systems and show how their performance can be improved with wise scheduling decisions.

Towards Optimization of Hybrid CPU/GPU Query Plans in Database Systems

Current database research identified the computational power of GPUs as a way to increase the performance of database systems. Since GPU algorithms are not necessarily faster than their CPU counterparts, it is important to use the GPU only if it is beneficial for query processing. In a general database context, only few research projects address hybrid query processing, i.e., using a mix of CPU- and GPU-based processing to achieve optimal performance. In this paper, we extend our CPU/GPU scheduling framework to support hybrid query processing in database systems. We point out fundamental problems and provide an algorithm to create a hybrid query plan for a query using our scheduling framework.

Database Scan Variants on Modern CPUs: A Performance Study

Main-memory databases rely on highly tuned database operations to achieve peak performance. Recently, it has been shown that different code optimizations for database operations favor different processors. However, it is still not clear how the combination of code optimizations (e.g., loop unrolling and vectorization) will affect the performance of database algorithms on different processors.

Load-aware inter-co-processor parallelism in database query processing

For a decade, the database community has been exploring graphics processing units and other co-processors to accelerate query processing. While the developed algorithms often outperform their CPU counterparts, it is not beneficial to keep processing devices idle while overutilizing others. Therefore, an approach is needed that efficiently distributes a workload on available (co-)processors while providing accurate performance estimates for the query optimizer. In this paper, we contribute heuristics that optimize query processing for response time and throughput simultaneously via inter-device parallelism. Our empirical evaluation reveals that the new approach achieves speedups up to 1.85 compared to state-of-the-art approaches while preserving accurate performance estimations. In a further series of experiments, we evaluate our approach on two new use cases: joining and sorting. Furthermore, we use a simulation to assess the performance of our approach for systems with multiple co-processors and derive some general rules that impact performance in those systems. Contribute heuristics to enhance performance by exploiting inter-device parallelismHeuristics consider load and speed on (co-)processors.Extensive evaluation on four use cases: aggregation, selection, sort, and joinAssess the performance of best heuristic for systems with multiple co-processorsDiscuss how operator-stream-based scheduling can be used in a query processor

Ocelot/HyPE: optimized data processing on heterogeneous hardware

The past years saw the emergence of highly heterogeneous server architectures that feature multiple accelerators in addition to the main processor. Efficiently exploiting these systems for data processing is a challenging research problem that comprises many facets, including how to find an optimal operator placement strategy, how to estimate runtime costs across different hardware architectures, and how to manage the code and maintenance blowup caused by having to support multiple architectures. In prior work, we already discussed solutions to some of these problems: First, we showed that specifying operators in a hardware-oblivious way can prevent code blowup while still maintaining competitive performance when supporting multiple architectures. Second, we presented learning cost functions and several heuristics to efficiently place operators across all available devices. In this demonstration, we provide further insights into this line of work by presenting our combined system Ocelot/HyPE. Our system integrates a hardware-oblivious data processing engine with a learning query optimizer for placement decisions, resulting in a highly adaptive DBMS that is specifically tailored towards heterogeneous hardware environments.

Generating Custom Code for Efficient Query Execution on Heterogeneous Processors

Processor manufacturers build increasingly specialized processors to mitigate the effects of the power wall in order to deliver improved performance. Currently, database engines have to be manually optimized for each processor which is a costly and error- prone process. In this paper, we propose concepts to adapt to and to exploit the performance enhancements of modern processors automatically. Our core idea is to create processor-specific code variants and to learn a well-performing code variant for each processor. These code variants leverage various parallelization strategies and apply both generic- and processor-specific code transformations. Our experimental results show that the performance of code variants may diverge up to two orders of magnitude. In order to achieve peak performance, we generate custom code for each processor. We show that our approach finds an efficient custom code variant for multi-core CPUs, GPUs, and MICs.