CS267: Applications of Parallel Computers

Shariq Rizvi

Research Interests

Platform and Coding Experiences

Misc

A database management system is a piece of software that allows storage, modification and querying of large amounts of data in an efficient manner. Research in this area has focused on efficient physical design of databases (disk-level storage issues like indexing), building independence between physical database design and the logical model available to the user (a key feature of the relational database model), efficient evaluation of queries (query processing and optimization) and other aspects of databases. Given the massive data and query workloads that database systems are expected to face, they form one of the most natural domains that can use parallelism.

Classification of Parallel Database Proposals

1. Shared memory: A collection of processors share the same physical memory. It is easiest to support as it can be run as a single-site DBMS and can depend on the OS for scheduling processes to multiple processors.


2. Shared disk: Each processor has its own memory but a shared set of storage disks is used. This presents several challenges rooted in the fact that consistency has to be preserved for the data residing on the disk under multiple processors accessing it. This become difficult as there is no shared memory.


3. Shared nothing: These involve multiple machines that share no hardware but communicate with each other using a high-speed LAN. Most research prototypes have been built based on this paradigm.

Case Study: Shared Nothing Architecture

• Storage: The relations (a fundamental notion in relational databases - represents a collection of "similar" tuples of data) are horizontally partitioned across all disk drives in the system. This parallelizes the scan operation, which reads all tuples from the relations. The system provides the user coarse-grained control over how the tuples of a relation can be partitioned across the disks (like round robin, hash based etc.).


• Indexing: The system allows both clustered and non-clustered indexes. On a user request, the index is created on each fragment of the relation.


• Query Processing: The query plan generator takes into account the partitioning scheme when scheduling the query over the different processor (e.g., a predicate like "key = a" in the query may require just the processor that has tuples corresponding to key value "a" to perform processing if hash based partitioning is used).

A Framework for Query Processing in Parallel Databases

The Volcano query processing system [3] is known for the extensibility that results from the uniform interface between operators. An operator is a query processing module that performs a specific function on the stream of data tuples that are sent to it. In general, a query execution plan can be expressed as a collection of operators and the way their output and input streams are connected to each other. Each operator in Volcano is coded such that it feels it is operating in a single-process environment. A special exchange operator encapsulates all the parallelism in the system and hence provides extensibility.

An overview of parallel database systems is provided in [4]. The authors argue that while in the early 80's, researchers felt that specialized hardware is key to large-scale scalable databases, around early 90's it was clear that parallel databases are a more natural software solution. They point out that relational operators very naturally extend to the parallel architecture as they operate on relations and output the same.

References