Building a Columnar Database on Shared Main Memory-Based Storage

نویسنده

  • Christian Tinnefeld
چکیده

In the field of disk-based parallel database management systems exists a great variety of solutions based on a shared-storage or a shared-nothing architecture. In contrast, main memory-based parallel database management systems are dominated solely by the shared-nothing approach as it preserves the in-memory performance advantage by processing data locally on each server. We argue that this unilateral development is going to cease due to the combination of the following three trends: a) Nowadays network technology features remote direct memory access (RDMA) and narrows the performance gap between accessing main memory inside a server and of a remote server to and even below a single order of magnitude. b) Modern storage systems scale gracefully, are elastic, and provide high-availability. c) A modern storage system such as Stanford’s RAMCloud even keeps all data resident in main memory. Exploiting these characteristics in the context of a main-memory parallel database management system is desirable. The advent of RDMA-enabled network technology makes the creation of a parallel main memory DBMS based on a shared-storage approach feasible. This thesis describes building a columnar database on shared main memory-based storage. The thesis discusses the resulting architecture (Part I), the implications on query processing (Part II), and presents an evaluation of the resulting solution in terms of performance, high-availability, and elasticity (Part III). In our architecture, we use Stanford’s RAMCloud as shared-storage, and the selfdesigned and developed in-memory AnalyticsDB as relational query processor on top. AnalyticsDB encapsulates data access and operator execution via an interface which allows seamless switching between local and remote main memory, while RAMCloud provides not only storage capacity, but also processing power. Combining both aspects allows pushing-down the execution of database operators into the storage system. We describe how the columnar data processed by AnalyticsDB is mapped to RAMCloud’s key-value data model and how the performance advantages of columnar data storage can be preserved. The combination of fast network technology and the possibility to execute database operators in the storage system opens the discussion for site selection. We construct a system model that allows the estimation of operator execution costs in terms of network transfer, data processed in memory, and wall time. This can be used for database operators that work on one relation at a time — such as a scan or materialize operation — to discuss the site selection problem (data pull vs. operator push). Since a database query translates to the execution of several database operators, it is possible that the optimal site selection varies per operator. For the execution of a database operator that works on two (or more) relations at a time, such as a join, the system model is enriched by additional factors such as the chosen algorithm (e.g. Gracevs. Distributed Block Nested Loop Join vs. Cyclo-Join), the data partitioning of the respective relations, and their overlapping as well as the allowed resource allocation. We present an evaluation on a cluster with 60 nodes where all nodes are connected via RDMA-enabled network equipment. We show that query processing performance is about 2.4x slower if everything is done via the data pull operator execution strategy (i.e. RAMCloud is being used only for data access) and about 27% slower if operator execution is also supported inside RAMCloud (in comparison to operating only on main memory inside a server without any network communication at all). The fast-crash recovery feature of RAMCloud can be leveraged to provide high-availability, e.g. a server crash during query execution only delays the query response for about one second. Our solution is elastic in a way that it can adapt to changing workloads a) within seconds, b) without interruption of the ongoing query processing, and c) without manual intervention. Titel der Arbeit Building a Columnar Database on Shared Main Memory-Based Storage: Database Operator Placement in a Shared Main Memory-Based Storage System that Supports Data Access and Code Execution Autor: Christian Tinnefeld Hasso-Plattner-Institut Betreuer: 1. Prof. Dr. Dr. h.c. Hasso Plattner, Hasso-Plattner-Institut 2. Prof. Dr. Donald Kossmann, ETH Zürich 3. Prof. John Ousterhout, PhD, Stanford University Zusammenfassung: Diese Arbeit beschreibt die Erstellung einer spalten-orientierten Datenbank auf einem geteilten, Hauptspeicher-basierenden Speichersystem. Motiviert wird diese Arbeit durch drei Faktoren. Erstens ist moderne Netzwerktechnologie mit “Remote Direct Memory Access” (RDMA) ausgestattet. Dies reduziert den Unterschied hinsichtlich Latenz und Durchsatz zwischen dem Speicherzugriff innerhalb eines Rechners und auf einen entfernten Rechner auf eine Größenordnung. Zweitens skalieren moderne Speichersysteme, sind elastisch und hochverfügbar. Drittens hält ein modernes Speichersystem wie Stanford’s RAMCloud alle Daten im Hauptspeicher vor. Diese Eigenschaften im Kontext einer spalten-orientierten Datenbank zu nutzen ist erstrebenswert. Die Arbeit ist in drei Teile untergliedert. Der erste Teile beschreibt die Architektur einer spalten-orientierten Datenbank auf einem geteilten, Hauptspeicher-basierenden Speichersystem. Hierbei werden die im Rahmen dieser Arbeit entworfene und entwickelte Datenbank AnalyticsDB sowie Stanford’s RAMCloud verwendet. Die Architektur beschreibt wie Datenzugriff und Operatorausführung gekapselt werden um nahtlos zwischen lokalem und entfernten Hauptspeicher wechseln zu können. Weiterhin wird die Ablage der nach einem relationalen Schema formatierten Daten von AnalyticsDB in RAMCloud behandelt, welches mit einem Schlüssel-Wertpaar Datenmodell operiert. Der zweite Teil fokussiert auf die Implikationen bei der Abarbeitung von Datenbankanfragen. Hier steht die Diskussion im Vordergrund wo (entweder in AnalyticsDB oder in RAMCloud) und mit welcher Parametrisierung einzelne Datenbankoperationen ausgeführt werden. Dafür werden passende Kostenmodelle vorgestellt, welche die Abbildung von Datenbankoperationen ermöglichen, die auf einer oder mehreren Relationen arbeiten. Der dritte Teil der Arbeit präsentiert eine Evaluierung auf einem Verbund von 60 Rechnern hinsichtlich der Leistungsfähigkeit, der Hochverfügbarkeit und der Elastizität vom System.

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

HyPer: Adapting Columnar Main-Memory Data Management for Transactional AND Query Processing

Traditionally, business applications have separated their data into an OLTP data store for high throughput transaction processing and a data warehouse for complex query processing. This separation bears severe maintenance and data consistency disadvantages. Two emerging hardware trends allow the consolidation of the two disparate workloads onto the same database state on one system: the increas...

متن کامل

A Cost-Aware and Workload-Based Index Advisor for Columnar In-Memory Databases

Optimal index configurations for in-memory databases differ significantly from configurations for their traditional disk-based counterparts. Operations such as full column scans that have previously been prohibitively expensive in disk-based and row-oriented databases are now computationally feasible with columnar main memory-resident data structures and even outperform index-based accesses in ...

متن کامل

Aggregates Caching in Columnar In-Memory Databases

The mixed database workloads found in enterprise applications are comprised of short-running transactional as well as analytical queries with resource-intensive data aggregations. In this context, caching the query results of long-running queries is desirable as it increases the overall performance. However, traditional caching approaches are inefficient in a way that changes in the base data r...

متن کامل

DataBlitz: A High Performance Main-Memory Storage Manager

DataBlitz’ is a main-memory storage management toolkit that supports the development of highperformance and fault-resilient applications requiring concurrent access to shared data. By combining transactions, data organization, and fault-resilience with direct access to data, DataBlitz bridges the gap between the speed of data structures in shared memory and the safety and convenience of a tradi...

متن کامل

Highspeed Graph Processing Exploiting Main-Memory Column Stores

A popular belief in the graph database community is that relational database management systems are generally ill-suited for efficient graph processing. This might apply for analytic graph queries performing iterative computations on the graph, but does not necessarily hold true for short-running, OLTP-style graph queries. In this paper we argue that, instead of extending a graph database manag...

متن کامل

ذخیره در منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

عنوان ژورنال:

دوره   شماره 

صفحات  -

تاریخ انتشار 2015