Question

Comprehensive System Design -advanced databases-

GPUs are new parallel hardware that encapsulates a large number of computing cores in a single device, therefore, providing massive parallel computing capabilities with low-cost. More information about the architecture of modern GPUs can be found in the attached paper SSDBM13.pdf. Given such details (as well as anything you can find online about GPUs), you are required to write an essay on the design of a DBMS that processes queries on GPUs. In your essay, you should address the following issues:

1. What types of database workloads are suitable for GPUs? Why? Hint: you can consider the following workload types: Online transaction processing (OLTP), online analytical processing (OLAP), and real-time data stream processing (DSP). You can search the Internet for more details about such workloads.

2. For the appropriate workload you identified in 1, sketch a DBMS design, and discuss the major components of such a DBMS, and how they are different from a traditional CPU-based relational DBMS;

3. What are the main limitations of GPUs in processing database workloads? If you have the freedom to design next generation of GPUs, what hardware/ software features would you introduce to speed up database queries?

Answer 1

1. Answer

• GPU supports for data base analytics involving online transaction processing (OLTP) and online analytical processing (OLAP) as these involves GPU databases giving results to complex queries in milliseconds there by speeding up data analytics.
• AI workloads such as Machine Learning(ML) algorithms working on parallel calculations of billions of data are suitable for GPU as they speed up calculations with high performance processing.
• GPU supports deep learning very well because the deep learning runs computing loops repeatedly on huge datasets running in gigabytes.
• Real time Data stream Processing (DSP ) needs GPU as it involves streaming of real time huge volumes of data like video and images. Because billions of multiple computations needed for three dimensional imaging , which can be taken care of only by GPU.
• GPU is much needed for big data analytics using No SQL databases like hadoop instead of traditional relational DBMS

2. Answer

• Oracle has introduced “Oracle Database In- memory “ using in-memory architecture supporting in-memory column format for OLAP along with traditional row formats for OLTP
• Oracle Database In- memory’s sql query optimizer sends the analytics OLAP queries to column format and OLTP queries to row format
• Oracle Database In- memory accelerates OLTP transactions by removing analytical indexes.
• Oracle uses new SIMD vector algorithms based on parallel stream paradigm which provide huge analytical performance.
• Traditional CPU based relational DBMS stored data in row format, which means each record of an entity is stored as a row in a database table which is not for analytics for faster retrieval of data but works good for online transactions systems.
• GPU based DBMS uses column format in which each attribute belonging to a record is stored in column structure which helps more for analytics for retrieving data faster.

3. Answer

a) Limitations

• GPU depend too much on floating point process mechanism. So the processing power is not sufficient to speed up database queries or its algorithms.
• GPU systems operates with PCI bus. It has its own high bandwidth local memory with huge bandwidth instead of reaching out to server memory. So GPU will process data by moving to and forth between PCI bus and other CPU.

b) Next generation GPU Design

• For next generation CPU design , we have to improvise existing software and hardware features by removing the existing bottlenecks in GPU and GPU based databases to support the queries speed up and in terms of performance for big data analytics, deep learning involving ML based workloads doing more fool proof researches on it.