Citation Context

...at in DryadLINQ, partly because the underlying MapReduce execution platform is much less flexible than Dryad. DryadLINQ and systems such as MapReduce are also distinguished from traditional databases =-=[25]-=- by having virtualized expression plans. The planner allocates resources independent of the actual cluster used for execution. This means both that DryadLINQ can run plans requiring many more steps th...

Citation Context

...o allow the vertex code to be automatically parallelized on a multi-core cluster computer. 4.6 Query Optimizations The DryadLINQ optimizer is similar in many respects to classical database optimizers =-=[12]-=-. It has a static component, which performs conditional graph rewrite rules on the EPG, and a dynamic component, which uses Dryad’s stage callbacks to optimize the execution at run time. Many of the c...

Citation Context

...S relies on the native filesystem for portability, many well-known databases can be configured to directly manage storage as raw disks at the application level, bypassing the filesystem entirely [1], =-=[12]-=-, [16]. Using storage in this manner allows the filesystem page cache to be bypassed in favor of an application cache, which eliminates double-buffering of data. Further, circumventing the filesystem ...

Citation Context

...but has remained an elusive goal, ―up in the clouds‖, for two decades. One can indeed easily separate the query processing and optimization components from the storage engine. However, as observed in =-=[10]-=-, ―The truly monolithic piece of a This article is published under a Creative Commons License Agreement (http://creativecommons.org/licenses/by/3.0/). You may copy, distribute, display, and perform th...

Citation Context

... management, and its subfield of concurrency control can be found in the textbooks by Bernstein et al. [2] and Weikum and Vossen [20]. The internal design of a DBMS is described by Hellerstein et al. =-=[9]-=-. The usual approach is for the lower, storage-focused parts of the DBMS to contain a component called a lock manager. Code that fetches or updates a record obtains an appropriate lock in the lock man...

Citation Context

...d a single data storage layout. This is a fixed decision that drives the whole design of the architecture of a database system. For example, traditional row-store systems store data one row at a time =-=[20]-=- while modern column-store systems store data one column at a time [1]. However, none of those choices is a universally good solution; different workloads require different storage layouts and data ac...

Citation Context

...tandard. The phantom problem, and the distinction between Repeatable Read and Serializable is described in [8]. The state-of-practice concurrency control for DBMS engines is presented very clearly in =-=[15]-=-. For locking approaches, the widely-used techniques include an efficient lock manager [13], with phantoms prevented by some form of next-key or index lock [19, 20, 18]. The alternative multiversion a...

Citation Context

...tabase management systems are notoriously monolithic pieces of software, largely because they use complex concurrency control mechanisms to orchestrate the simultaneous execution of many transactions =-=[9]-=-. Many attempts have been made—with varying success—to build clean interfaces between various components, decoupling transaction coordination, buffer pool management, logging/recovery mechanisms, data...

Citation Context

... main problem of the greedy algorithm introduced in Section 5.1 is that it tends to pack lots of tenants on a server when the tenants are small. And in multitenant databases, tenants are indeed small =-=[26, 18]-=-. Therefore, the idea of improvement is to proactively create new servers, even if it increases the current expected cost. To achieve this, we introduce the concept of half full servers. This is used ...