Introducing PgOpenCL

A New PostgreSQL
Procedural Language
Unlocking the Power of the GPU!
By
Tim Child
 Bio

Tim Child
• 35 years experience of software development
• Formerly
• VP Oracle Corporation
• VP BEA Systems Inc.
• VP Informix
• Leader at Illustra, Autodesk, Navteq, Intuit, …
• 30+ years experience in 3D, CAD, GIS and DBMS
 Terminology
Term Description
Procedure Language Language for SQL Procedures (e.g. PgPLSQL, Perl, TCL, Java, … )
GPU Graphics Processing Unit (highly specialized CPU for graphics)
GPGPU General Purpose GPU (non-graphics programming on a GPU)
CUDA Nvidia’s GPU programming environment
APU Accelerated Processing Unit (AMD’s Hybrid CPU & GPU chip)
ISO C99 Modern standard version of the C language
OpenCL Open Compute Language
OpenMP Open Multi-Processing (parallelizing compilers)
SIMD Single Instruction Multiple Data (Vector instructions )
SSE x86, x64 (Intel, AMD) Streaming SIMD Extensions
xPU Any Processing Unit device (CPU, GPU, APU)
Kernel Functions that execute on a OpenCL Device
Work Item Instance of a Kernel
Workgroup A group of Work Items
FLOP Floating Point Operation (single = SQL real type )
MIC Many Integrated Cores (Intel’s 50+ x86 Core chip architecture)
 Some Technology Trends
Impacting DBMS
• Solid State Storage
– Reduced Access Time, Lower Power, Increasing in capacity
• Virtualization
– Server consolidation, Specialized VM’s, lowers direct costs
• Cloud Computing
– EC2, Azure, … lowers capital requirements
• Multi-Core
– 2,4,6,8, 12, …. Lots of benefits to multi-threaded applications

• xPU (GPU/APU)
– GPU >1000 Cores
– > 1T FLOP /s @ €2500
– APU = CPU + GPU Chip Hybrids due in Mid 2011
– 2 T FLOP /s for $2.10 per hour (AWS EC2)
– Intel MIC “Knights Corner “ > 50 x86 Cores
 Compute Intensive
xPU Database Applications
• Bioinformatics

• Signal/Audio/Image Processing/Video

• Data Mining & Analytics

• Searching

• Sorting

• Spatial Selections and Joins

• Map/Reduce

• Scientific Computing

• Many Others …
 GPU vs CPU
Vendor NVidia ATI Radeon Intel
Architecture Fermi Evergreen Nehalem
Cores 448 1600 4
Simple Simple Complex
Transistors 3.1 B 2.15 B 731 M
Clock 1.5 G Hz 851 M Hz 3 G Hz
Peak Float 1500 G 2720 G 96 G
Performance FLOP / s FLOP / s FLOP / s
Peak Double 750 G 544 G 48 G
Performance FLOP / s FLOP / s FLOP / s
Memory ~ 190 G / s ~ 153 G / s ~ 30 G / s
Bandwidth
Power 250 W > 250 W 80 W
Consumption
SIMD / Vector Many Many SSE4+
Instructions
 Multi-Core Performance

Source NVidia
 Future (Mid 2011)
APU Based PC
APU (Accelerated Processing Unit)

APU Chip
CPU CPU ~20 GB/s System RAM

North Bridge
~20 GB/s APU’s
PCIE ~12 GB/s

Adds an Embedded
Embedded GPU
GPU

Discrete
150 GB/s Graphic RAM
GPU

Source AMD
 Scalar vs. SIMD
Scalar Instruction
C=A+B 1 + 2 = 3

SIMD Instruction 1 3 5 7

+
Vector C = Vector A + Vector B 2 4 6 8

=
3 7 11 15

OpenCL
Vector lengths 2,4,8,16 for char, short, int, float, double
 Summarizing xPU
Trends
• Many more xPU Cores in our Future
• Compute Environment becoming Hybrid
– CPU and GPU’s
– Need CPU to give access to GPU power
• GPU Capabilities
– Lots of cores
– Vector/SIMD Instructions
– Fast Memory
• GPU Futures
– Virtual Memory
– Multi-tasking / Pre-emption
 Scaling PostgreSQL Queries
on xPU’s
Multi-Core CPU Many Core GPU

PgOpenCL PgOpenCL PgOpenCL PgOpenCL PgOpenCL PgOpenCL PgOpenCL PgOpenCL PgOpenCL

Threads Threads Threads Threads Threads Threads Threads Thread Thread

PgOpenCL PgOpenCL PgOpenCL PgOpenCL PgOpenCL

Postgres Threads Threads Threads Thread Thread
Process

PgOpenCL PgOpenCL PgOpenCL PgOpenCL

PgOpenCL
Threads Threads Thread Thread
Threads

Using More
Transistors
 Parallel
Programming Systems
Category CUDA OpenMP OpenCL
Language C C, Fortran C
Cross Platform X √ √
Standard Vendor OpenMP Khronos
CPU X √ √
GPU √ X √
Clusters X √ X

Compilation / Link Static Static Dynamic

 What is OpenCL?
• OpenCL - Open Compute Language
– Subset of C 99
– Open Specification
– Proposed by Apple
– Many Companies Collaborated on the Specification
– Portable, Device Agnostic
– Specification maintained by Khronos Group
• PgOpenCL
– OpenCL as a PostgreSQL Procedural Language
 System Overview
DBMS Server

PgOpenCL
PgOpenCL
Web HTTP Web SQL SQL
SQL
Browser Server Statement Procedure
Procedure

PCIe X2 Bus
TCP/IP

App
PostgreSQL GPGPU
Server

Disk I/O Tables

TCP/IP
PostgreSQL
Client
 OpenCL
Language
• A subset of ISO C99
– - But without some C99 features such as standard C99 headers,
– function pointers, recursion, variable length arrays, and bit fields
• A superset of ISO C99 with additions for:
– - Work-items and Workgroups
– - Vector types
– - Synchronization
– - Address space qualifiers
• Also includes a large set of built-in functions
– - Image manipulation
– - Work-item manipulation,
– - Specialized math routines, etc.
 PgOpenCL
Components
• New PostgreSQL Procedural Language
– Language handler
• Maps arguments
• Calls function
• Returns results
– Language validator
• Creates Function with parameter & syntax checking
• Compiles Function to a Binary format
• New data types
– cl_double4, cl_double8, ….
• System Admin Pseudo-Tables
– Platform, Device, Run-Time, …
PgOpenCL
Admin
 PGOpenCL
Function Declaration
CREATE or REPLACE FUNCTION VectorAdd(IN a float[], IN B float[], OUT c float[])
AS $BODY$

#pragma PGOPENCL Platform : ATI Stream

#pragma PGOPENCL Device : CPU

__kernel __attribute__((reqd_work_group_size(64, 1, 1)))

void VectorAdd( __global const float *a, __global const float *b, __global float *c)
{
int i = get_global_id(0);

c[i] = a[i] + b[i];

}

$BODY$
Language PgOpenCL;
 PgOpenCL
Execution Model
A
Table
B

Select Table 100’s - 1000’s of

to Array Threads (Kernels)

xPU
VectorAdd(A, B)
A + B Returns C = C

Copy Unnest Array

Copy To Table
Table

C C C C C C C C C C C C C
 Using
Re-Shaped Tables
100’s - 1000’s of
Table of Threads (Kernels) Table of
Arrays Arrays
A + B = C

A
C C C C
B
xPU
VectorAdd(A, B)
Returns C
A
C C C C
B

Copy
Copy
 Today’s GPGPU
Challenges
• No Pre-emptive Multi-Tasking
• No Virtual Memory
• Limited Bandwidth to discrete GPGPU
– 1 – 8 G/s over PCIe Bus
• Hard to Program
– New Parallel Algorithms and constructs
– “New” C language dialect
• Immature Tools
– Compilers, IDE, Debuggers, Profilers - early years
• Data organization really matters
– Types, Structure, and Alignment
– SQL needs to Shape the Data
• Profiling and Debugging is not easy

Solves Well for Problem Sets with the Right Shape!

 Making a Problem
Work for You
• Determine % Parallelism Possible
for ( i = 0, i < ∞, i++)
for ( j = 0; j < ∞; j++ )
for ( k = 0; k < ∞; k++ )

• Arrange data to fit available GPU RAM

• Ensure calculation time >> I/O transfer overhead
• Learn about Parallel Algorithms and the OpenCL language
• Learn new tools
• Carefully choose Data Types, Organization and Alignments
• Profile and Measure at Every Stage
 PgOpenCL
System Requirements
• PostgreSQL 9.x
• For GPU’s
– AMD ATI OpenCL Stream SDK 2.x
– NVidia CUDA 3.x SDK
– Recent Macs with O/S 11.6
• For CPU’s (Pentium M or more recent)
– AMD ATI OpenCL Stream SDK 2.x
– Intel OpenCL SDK Alpha Release (x86)
– Recent Macs with O/S 11.6
 PGOpenCL
Status
Today 1Q 2011
Prototype Beta

2010 2011

• Wish List
• Beta Testers
– Existing OpenCL App?
– Have a GPU App?
• Contributors
– Code server side functions?
• Sponsors & Supporters
– AMD Fusion Fund?
– Khronos?
 PgOpenCL
Future Plans
• Increase Platform Support
• Scatter/Gather Functions
• Additional Type Support
– Image Types
– Sparse Matrices
• Run-Time
– Asynchronous
– Events
– Profiling
– Debugging
 Using the
Whole Brain
APU Chip
PgOpenCl PgOpenCl
PgOpenCL PgOpenCL
CPU
CPU CPU
Postgres You can’t be in a
parallel universe
with a single
brain!
North Bridge
~20 GB/s
• Heterogeneous Compute Environments
PgOpenCl
PgOpenCl • CPU’s, GPU’s, APU’s
Embedded PgOpenCl • Expect 100’s – 1000’s of cores
PgOpenCl
GPU PgOpenCL

The Future Is Parallel: What's a Programmer to Do?

 Summarizing
PgOpenCL
• Supports Heterogeneous Parallel Compute Environments
• CPU’s, GPU’s, APU’s

• OpenCL
• Portable and high-performance framework
–Ideal for computationally intensive algorithms
–Access to all compute resources (CPU, APU, GPU)
–Well-defined computation/memory model
•Efficient parallel programming language
–C99 with extensions for task and data parallelism
–Rich set of built-in functions
•Open standard for heterogeneous parallel computing
• PgOpenCL
• Integrates PostgreSQL with OpenCL
• Provides Easy SQL Access to xPU’s
• APU, CPU, GPGPU
• Integrates OpenCL
• SQL + Web Apps(PHP, Ruby, … )
 More
Information
• PGOpenCL
• Twitter @3DMashUp

• OpenCL

• www.khronos.org/opencl/

• www.amd.com/us/products/technologies/stream-technology/opencl/

• http://software.intel.com/en-us/articles/intel-opencl-sdk

• http://www.nvidia.com/object/cuda_opencl_new.html

• http://developer.apple.com/technologies/mac/snowleopard/opencl.html
 Q&A

• Using Parallel Applications?

• Benefits of OpenCL / PgOpenCL?
• Want to Collaborate on PgOpenCL?