Felix Beier

Efficient co-processor utilization in database query processing

Specialized processing units such as GPUs or FPGAs provide great opportunities to speed up database operations by exploiting parallelism and relieving the CPU. However, distributing a workload on suitable (co-)processors is a challenging task, because of the heterogeneous nature of a hybrid processor/co-processor system. In this paper, we present a framework that automatically learns and adapts execution models for arbitrary algorithms on any (co-)processor. Our physical optimizer uses the execution models to distribute a workload of database operators on available (co-)processing devices. We demonstrate its applicability for two common use cases in modern database systems. Additionally, we contribute an overview of GPU-co-processing approaches, an in-depth discussion of our frameworks operator model, the required steps for deploying our framework in practice and the support of complex operators requiring multi-dimensional learning strategies.

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.

GiST scan acceleration using coprocessors

Efficient lookups in huge, possibly multi-dimensional datasets are crucial for the performance of numerous use cases that generate multiple search operations at the same time, like point queries in ray tracing or spatial joins in collision detection of interactive 3D applications. These applications greatly benefit from index structures that quickly filter relevant candidates for further processing. Since different lookup operations are independent from each other, they might be processed in parallel on modern hardware like multi-core CPUs or GPUs. But implementing efficient algorithms for all kinds of indexes on various hardware platforms is a challenging task. In this paper, we present a new approach that extends the existing GiST index framework with an abstraction layer for the hardware where index operations are executed. Furthermore, we provide first performance evaluations for the scan execution on CPUs and an Nvidia Tesla GPU.

Online reorganization in read optimized MMDBS

Query performance is a critical factor in modern business intelligence and data warehouse systems. An increasing number of companies uses detailed analyses for conducting daily business and supporting management decisions. Thus, several techniques have been developed for achieving near realtime response times - techniques which try to alleviate I/O bottlenecks while increasing the throughputs of available processing units, i.e. by keeping relevant data in compressed main-memory data structures and exploiting the read-only characteristics of analytical workloads. However, update processing and skews in data distribution result in degenerations in these densely packed and highly compressed data structures affecting the memory efficiency and query performance negatively. Reorganization tasks can repair these data structures, but -- since these are usually costly operations -- require a well-considered decision which of several possible strategies should be processed and when, in order to reduce system downtimes. In this paper, we address these problems by presenting an approach for online reorganization in main-memory database systems (MMDBS). Based on a discussion of necessary reorganization strategies in IBM Smart Analytics Optimizer, a read optimized parallel MMDBS, we introduce a framework for executing arbitrary reorganization tasks online, i.e. in the background of normal user workloads without disrupting query results or performance.

Integrating Cluster-Based Main-Memory Accelerators in Relational Data Warehouse Systems

Today, data warehouse systems are faced with challenges for providing nearly realtime response times even for complex analytical queries on enormous data volumes. Highly scalable computing clusters in combination with parallel in-memory processing of compressed data are valuable techniques to address these challenges. In this paper, we give an overview on core techniques of the IBM Smart Analytics Optimizer—an accelerator engine for IBM’s mainframe database system DB2 for z/OS. We particularly discuss aspects of a seamless integration between the two worlds and describe techniques exploiting features of modern hardware such as parallel processing, cache utilization, and SIMD. We describe issues encountered during the development and evaluation of our system and outline current research activities for solving them.

Architecture of a data analytics service in hybrid cloud environments

Abstract DB2 for z/OS is the backbone of many transactional systems in the world. IBM DB2 Analytics Accelerator (IDAA) is IBMs approach to enhance DB2 for z/OS with very fast processing of OLAP and analytical SQL workload. While IDAA was originally designed as an appliance to be connected directly to System z, the trend in the IT industry is towards cloud environments. That offers a broad range of tools for analytical data processing tasks. This article presents the architecture for offering a hybrid IDAA, which continues the seamless integration with DB2 for z/OS and now also runs as a specialty engine in cloud environments. Both approaches have their merit and will remain important for customers in the next years. The specific challenges for accelerating query processing for relational data in the cloud are highlighted. Specialized hardware options are not readily available, and that has a direct impact on the system architecture, the offered functionality and its implementation.

GPU-GIST – a case of generalized database indexing on modern hardware

Abstract A lot of different indexes have been developed for accelerating search operations on large data sets. Search trees, representing the most prominent class, are ubiquitous in database management systems but are also widely used in non-DBMS applications. An approach for lowering the implementation complexity of these structures are index frameworks like generalized search trees (GiST). Common data management operations are implemented within the framework which can be specialized by data organization and evaluation strategies in order to model the actual index type. These frameworks are particularly useful in scientific and engineering applications where characteristics of the underlying data set are not known a priori and a lot of prototyping is required in order to find suitable index structures for the workload. However, existing frameworks only abstract data organization and data maintenance aspects to model different index families, while traversal operations for executing searches are implemented serially. This paper presents an approach for enabling parallel processing in GiST in order to leverage the full power of parallel processor architectures for different index implementations at once. Further, results of a prototypical implementation are evaluated on a hybrid CPU/GPU system architecture to verify the applicability of this generic framework idea on different hardware platforms.

Autonomous workload-driven reorganization of column groupings in MMDBS

A current trend to achieve high query performance even for huge data warehouse and business intelligence systems is to exploit main-memory-based processing techniques such as compression, cache-conscious strategies, and optimized data structures. However, update processing and skews in data distribution might lead to degenerations in such densely packed and highly compressed data structures affecting the memory efficiency and query performance negatively. Thus, reorganization tasks for repairing these data structures are necessary but should be carefully applied in order to not impact query execution or even system availability significantly. In this paper, we consider the special problem of tuple layout in banked storage structures. Based on runtime statistics capturing typical access patterns in the current workload, we present a bank reassignment approach that can be piggybacked to maintenance tasks without any administration overhead. We have implemented this approach in IBM Smart Analytics Optimizer (ISAOPT). The results of our experimental evaluation show that a simple automatic restructuring of the considered hybrid row-column-store structures offers opportunities to improve query runtimes when a slight memory overhead is acceptable.