DATA-INTENSIVE COMPUTING IN SQL WITH GPUS

Tamás Budavári, Richard Wilton / Johns Hopkins U.10/7/2010

Tamás Budavári

10/7/2010

Basic Concept2

Implement computational functionality in a separate process from the database server process

Access through inter-process communication (IPC)

IPCSQL code

SQL-callable procedureCUDA functionality

Database server Out-of-process server

Tamás Budavári

10/7/2010

3

Why Do It This Way?

IPCSQL code

SQL-callable procedureCUDA functionality

Database server Out-of-process server

Avoid DBMS permissions, threading restrictions, etc.

Encapsulate native-code functionality (CUDA API calls)

Exploit lower-level APIs to facilitate data movement between DBMS and CUDA implementation

Tamás Budavári

10/7/2010

4

SQL Callable Procedure

IPCSQL code

SQL-callable procedureCUDA functionality

Database server Out-of-process server

SQLCLR procedure in C# Fills a shared-memory buffer with a SQL result set

Signals the out-of-process server and waits for a response

Returns data from the IPC buffer to the database server

Tamás Budavári5

10/7/2010

Out-Of-Process Server

IPCSQL code

SQL-callable procedureCUDA functionality

Database server Out-of-process server

Dynamically loads the CUDA implementation

Invokes a method with a reference to the IPC buffer

Signals the SQL-callable procedure that data is ready

Tamás Budavári

Case Study 1: Spatial Statistics

10/7/2010

Correlation functions

Histogram of distances

State of the art

Dual-tree traversal

High resolution bins?

Just like brute force

8 bins

6

Tamás Budavári

Case Study 1: Spatial Statistics

10/7/2010

Correlation functions

Histogram of distances

State of the art

Dual-tree traversal

High resolution bins?

Just like brute force

8 bins

7

800 × 800 bins

Tamás Budavári

User-Defined Functions

Pair counts computed on GPU

Returns 2D histogram as a table (i, j, cts)

Calculate the correlation fn in SQL

10/7/2010

8

Tamás Budavári

User-Defined Functions

Pair counts computed on GPU

Returns 2D histogram as a table (i, j, cts)

Calculate the correlation fn in SQL

10/7/2010

9

Tamás Budavári

Async SQL Interface

10/7/2010

10

Tamás Budavári

800 × 800 bins11

59

0 T

rill

ion

Gal

axy

Pai

rs800 × 800 bins

Tamás Budavári

Summary

10/7/2010

New Moore’s law in GPUs – let’s use them!

Novel extension to SQL Server

Provides direct access to multiple GPUs

Inter- and intra-task GPU parallelism

Also enables distributed parallel queries, MPI…

Scientific applications

Corr fn, xmatch, photo-z and gene sequencing…

12

Tamás Budavári

10/7/2010

13