PG-Strom

PG-Strom
GPU Accelerated Asynchronous
Super-Parallel Query Execution

KaiGai Kohei <kaigai@kaigai.gr.jp>
(Tw: @kkaigai)

Self Introduction

• 名前海外浩平 (KaiGai Kohei)
• 所属 NEC Europe - SAP Global Competence Center
• 仕事 OSS活用によるイノベーション創出 (20-30%)
SAPとのアライアンス、PF製品の拡販 (70-80%)
• SAPのIn-memory DB “SAP HANA” の認証作業とか
• CLUSTERPROのSAP認定取得、拡販とか特にコレの
割合がデカい

20130218 - PG-Strom Workshop, Tokyo 2

All everyone talks about BIG-DATA

猫杓子

熱い視線

BIG DATA

Big-Data Database?

¥¥¥¥¥¥
$$$$$$
€€€€€€

Homogeneous / Heterogeneous computing

KPIs
Homogeneous
Scale-Up
• Computing Performance
• Power Consumption
• System Cost (HW/SW)
Heterogeneous • Variety of Applications
Scale-Up
• Vendor Support
• Software Development
:

+
Scale-out
(not a topic of
today’s talk)


Design concept of PG-Strom

The world cheapest
The most Cost-Effective
Big-Data Database
• Utilization of open source technology
• Utilization of commodity hardware
• up-to ?? CPUs
まだ、この辺をとやかく
• up-to ??? GB RAM 言える段階ではない
• up-to ??? Data size
• Utilization of heterogeneous computing with GPU

Characteristics of GPU (1/2)

Nvidia AMD Intel
Kepler GCN SandyBridge

Model Tesla K20X FirePro S9000 Xeon E5-2690
(Q4/2012) (Q3/2012) (Q1/2012)
Number of 7.1billion 4.3billion 2.26billion
Transistors
Number of 2688 1792 16
Cores Simple Simple Functional
Core clock 732MHz 925MHz 2.9GHz
Peak FLOPS 3.95Tflops 3.23TFlops 185.6GFlops
Memory 6GB, GDDR5 6GB, GDDR5 384GB/socket,
Size / TYPE DDR3
Memory ~192GB/s ~264GB/s ~51.2GB/s
Bandwidth
Power ~235W ~225W ~135W
Consumption
Price $3199? $2499? $2061


Characteristics of GPU (2/2)

Example)

Zi = Xi + Yi (0 <= i <= n)

X0 X1 X2 Xn

+ + + +
Y0 Y1 Y2 Yn

   
Z0 Z1 Z2 Zn

Assign a particular “core”

Nvidia’s GeForce GTX 680 Block Diagram

Play with GPU (1/3)
Memory CPU GPU
計算負荷
GPUの仕事
GPGPU
on-host DDR3-1600
(non-integrated)
buffer (~51.2GB/s)

DDR5
通常のI/O負荷 192.2GB/s
IO HUB
 CPUの仕事 GPU
on-device
code
HDD buffer
HBA device DRAM

SAS 2.0 (600MB/s) PCI-E 3.0 x16 (~32GB/s)

 Asynchronous Execution of CPU, GPU and PCI-E
 Minimization of data transfer between host and device

Play with GPU (2/3)
Host code example
void sqrt_float4(int n, float v[])
{
/* Acquire device memory and data transfer (host -> device) */
dev_v = clCreateBuffer(cxt, CL_MEM_READ_WRITE,
sizeof(float) * n, NULL, &rv);
/* Enqueue data transfer host to device */
clEnqueueWriteBuffer(cmdq, dev_x, CL_TRUE, 0, NULL,
v, 0, NULL, NULL);
/* Set arguments of kernel code */
clSetKernelArg(kernel, 0, sizeof(int), (void *)&n);
clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&dev_v);
/* Enqueue invocation of device kernel */
clEnqueueNDRangeKernel(cmdq, kernel, 1, NULL, &g_itemsz, &l_itemsz,
0, NULL, NULL);
/* Enqueue data transfer device to host */
clEnqueueReadBuffer(cmdq, dev_x, CL_TRUE, 0, NULL,
v, 0, NULL, NULL);
/* Release device memory */
clReleaseMemObject(dev_x)
}


Play with GPU (3/3)
Device code example
__kernel void dev_sqrt_float(int length, float x[])
{
int i = get_global_id(0);

if (i < length)
x[i] = sqrt(x[i]);
}

Host code to load kernel
/* Load source code of the program */
program = clCreateProgramWithSource(cxt, 1,
(const char *)&kernel_source,
(const size_t *)&kernel_source_len, &rv);
/* Run-time build of the program */
rv = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL);
/* Create a device kernel object */
kernel = clCreateKernel(program, “dev_sqrt_float”, &rv);


Comparison - CPU vs CPU + GPU

CPU CPU + GPU • Advantage
Storage  Storage 
• シンプルな演算の
Host buffer Host buffer 超並列処理
• 非同期実行＆
DMA: host  device
パイプライン処理
Storage  Parallel • Disadvantage
Host buffer Calculation
• ホストデバイス間の
Loop & DMA転送コスト
DMA: host  device
Calculation
Host buffer • コードの複雑化
 output

depends on
workload
Host buffer
 output


Architecture
of
PG-Strom


PG-Strom’s Asynchronous Execution model
vanilla PostgreSQL PostgreSQL with PG-Strom

CPU CPU GPU

Asynchronous memory
transfer and execution

Iteration of scan tuples and
evaluation of qualifiers

Synchronization

Larger “chunk” to scan
Earlier than
the database at once
“Only CPU” scan

: Scan tuples on shared-buffers
: Execution of the qualifiers
Page
20130218 - PG-Strom Workshop, Tokyo
14

Re-definition of SQL/MED (1/2)

• SQL/MED (Management of External Data)
• External data source performing as if regular tables
• Not only “management”, but external computing resources also
Exec
Regular Exec
storage
Query Executor Table
Query Planner
Query Parser

Foreign MySQL
Table FDW

SQL
Query Foreign Oracle
FDW Exec
Table

Foreign PG-Strom
Table FDW

Exec
Regular
storage
Table


Re-definition of SQL/MED (2/2)
Query
Parser
SQL/MED API construction of
Query remote SQL
Tree remote
SQL
Query FDW
Planner Planner

Remote pgsql
connection open

remote query
Query FDW
Executor Executor
result set

Result connection close
Set

Extension module 16

PG-Strom as SQL/MED driver
Query
Parser ① 条件句から、GPU用
Kernel Codeを自動生成
Query
Tree WHERE log(x) < 10

.....
Query FDW ....
Planner Planner .....

② Kernel codeの ④ GPU Kernelの
JIT-compile 非同期実行

chunk
Query FDW
buffer
Executor Executor

③ Shadow Table
Result
からのロード
Set

PG-Strom module shadow tables
17

Overall architecture
World of CPU
regular shadow
tables tables

shared_buffer chunk_buffer World of GPU
chunk GPU
data code
GPU Device
chunk Memory
SeqScan, PG-Strom request data
etc... handler JIT Super
compile Parallel
ForeignScan
Event Execution
Result monitor GPU Kernel
Query Executor GPU Function
PG-Strom kernel
PostgreSQL Backend GPU Server

Postmaster
background worker
18

So what, How fast is it?
postgres=# SELECT COUNT(*) FROM rtbl
WHERE sqrt((x-256)^2 + (y-128)^2) < 40;
count
--------
100467
(1 row)
Time: 7668.684 ms
postgres=# SELECT COUNT(*) FROM ftbl
WHERE sqrt((x-256)^2 + (y-128)^2) < 40;
count
--------
100467
(1 row) Accelerated!
Time: 857.298 ms
 CPU: Xeon E5-2670 (2.60GHz), GPU: NVidia GeForce GT640, RAM: 384GB
 Both of regular rtbl and PG-Strom ftbl contain 20milion rows with same value

Key Technologies

• Automatic GPU code generation & JIT compile
• Column-oriented data structure
• Asynchronous Execution


Automatic “pseudo” code generation

SELECT * FROM ftbl WHERE
c like ‘%xyz%’ AND sqrt((x-256)^2+(y-100)^2) < 10;

contains unsupported
operators / functions Translation to
pseudo code
xreg10 = $(ftbl.x)
xreg12 = 256.000000::double
Pseudo-code based implementation
will be replaced by native code and xreg8 = (xreg10 - xreg12)
JIT-compile approach soon. xreg10 = 2.000000::double
xreg6 = pow(xreg8, xreg10)
xreg12 = $(ftbl.y)
xreg14 = 128.000000::double
:


Automatic native code generation - WIP

SELECT * FROM ftbl WHERE
c like ‘%xyz%’ AND sqrt((x-256)^2+(y-100)^2) < 10;

OpenCL run-time builds
native GPU binary
__kernel void
pgstrom_qual(int nitems, bool result[],
float x[], float y[])
{
int index = get_global_id(0);

if (sqrt(pow(x[i] - 256, 2) +
pow(y[i] - 100, 2)) < 10)
result[i] = true;
else
result[i] = false;
}


Save bandwidth & shared-buffer usage
E.g) SELECT name, tel, email, address FROM address_book
WHERE sqrt((pos_x – 24.5)^2 + (pos_y – 52.3)^2) < 10;
 No sense to fetch columns being not in use
CPU GPU CPU GPU

Synchronization Synchronization

: Scan tuples on the shared-buffers
 Save the bandwidth of PCI-E bus
: Execution of the qualifiers
: Columns being not used the qualifiers  Save the shared-buffer usage

Column-oriented data structure (1/3)

(shadow) TABLE “public.ft.rowid”
rowid nitems isnull
FOREIGN TABLE ft
4000 2000 {0,0,0,1,0,0,…}
int float text
X Y Z 6000 2000 {0,0,0,0,0,0,…}
: : :
(shadow) TABLE “public.ft.z.cs”
14000 400 {0,0,1,0,0,0,…}
rowid nitems isnull values
4000 15 {0,0,…} { ‘hello’, ‘world’, … }
4015 20 {0,0,…} { ‘aaa’, ‘bbb’, ‘ccc’, … }
(shadow) TABLE “public.ft.y.cs”
rowid nitems : isnull : : values :
4000 25014275
{0,0,…} { 1.38, 6.45, 2.15, …‘yyy’, ‘zzz’, …}
25 {0,0,…} {‘xxx’, }
4250 250 {0,1,…} { 4.32, 5.46, 3.14, … }
(shadow): TABLE “public.ft.a.cs”
: : :
rowid nitems 14200
isnull 100 values
{0,0,…} {19, 29, 39, 49, 59, …}
4000 500 {0,0,…} {10, 20, 30, 40, 50, …}
4500 500 {0,1,…} {11, 0, 31, 41, 51, …}
: : : :
14200 200 {0,0,…} {19, 29, 39, 49, 59, …}


postgres=# CREATE FOREIGN TABLE example
(a int, b text) SERVER pg_strom;
CREATE FOREIGN TABLE

postgres=# SELECT * FROM pgstrom_shadow_relations;
oid | relname | relkind | relsize
-------+----------------------+---------+-----------
16446 | public.example.rowid | r | 0
16449 | public.example.idx | i | 8192
16450 | public.example.a.cs | r | 0
16453 | public.example.a.idx | i | 8192
16454 | public.example.b.cs | r | 0
16457 | public.example.b.idx | i | 8192
16462 | public.example.seq | S | 8192
(9 rows)

postgres=# SELECT * FROM pg_strom."public.example.a.cs" ;
rowid | nitems | isnull | values
-------+--------+--------+--------
(0 rows)


② Calculation opcode Pseudo Code

PgStromChunkBuffer
① Transfer rowmap

value a[] <not used>

value b[]
③ Write-Back
value c[]

value d[] <not used>

• Less bandwidth consumption Table: my_schema.ft1.b.cs
of PCI-Express bus 10100 {2.4, 5.6, 4.95, … }
10300 {10.23, 7.54, 5.43, … }
• Less usage of buffer-cache
Table: my_schema.ft1.c.cs
• Suitable for data 10100 {‘2010-10-21’, …}
compression 10200 {‘2011-01-23’, …}
10300 {‘2011-08-17’, …}

PGconf.EU 2012 / PGStrom - GPU Accelerated Asynchronous Execution Module 26

Asynchronous Execution (1/2)
IterateForeignScan
Yes free_chunk_list
No more rows
on current chunk?
current
chunk Job Queue
No

next
If no chunks are ready yet

Load chunk from
chunk
shadow tables

GPU
code

shadow tables

GPU Management
current Server
Return next
chunk
TupleTableSlot
ready_chunk_list

Asynchronous Execution (2/2) - in the future
IterateForeignScan
Yes free_chunk_list
Asynchronous Asynchronous
No more rows Data Load Calculation
on current chunk?
current
chunk
No
next Job Queue

Job Queue
chunk
next
chunk

GPU shadow tables
code
Parallel I/O Server
Parallel I/O Server
Parallel I/O Server

GPU Management
current Server
Return next
chunk
TupleTableSlot
ready_chunk_list

Eco-System
in
PostgreSQL
Development


PostgreSQL developer’s community

PostgreSQL
developer’s
community

contribution, software,
feedback, documentation,
donation, ... knowledge, ...
Service
infrastructure,
support,
consulting, ... 30

PostgreSQL development cycle
2011 2012 2013
v9.2
cycle
v9.2 Release
CommitFest 1st~4th
v9.3 PGconf2011 &
cycle developer meeting
CommitFest 1st~4th
PGconf2012 &
developer meeting

v9.3 development schedule
• 17th-May developer meeting
• 15th-Jun CommitFest:1st
• 15th-Sep CommitFest:2nd
• 15th-Nov CommitFest:3rd
PostgreSQL developer meeting (17th-May-2012)
• 15th-Jan CommitFest:4th

PostgreSQL CommitFest


Key features towards upcoming v9.3 (1/3)

• Background Worker
• It enables extensions to manage own background worker process
• Pre-requisite of PG-Strom’s GPU control server
• KaiGai implemented 1st version, then Alvaro revised and committed

Shared Resources (DB cluster, shared mem, IPC, ...)

Built-in Extra
background daemon
PostgreSQL
PostgreSQL
PostgreSQL
PostgreSQL
Backend
Backend
Backend (autovacuum, Own main()
Backend bgwriter...)

manage
Extension
postmaster



• Writable-FDW
• It allows FDW-drivers to modify external data source via foreign-table.
• Several new APIs shall be added
• Helpful for PG-Strom to modify shadow-tables using standard DML
• KaiGai submitted patch, then it is “ready-for-committer” status now

SELECT
SQL Executor
SQL Planner
SQL Parser

INSERT
FDW
driver
UPDATE

DELETE
New API External
Data Source



• Writable-FDW (Pseudo-column support)
• It required to identify a particular remote-row to be written.
• “rowid” shall be carried from scan-stage to modify-stage as
a value of pseudo-column.
• Pseudo-column concept is also available to push-down complex
calculation into external computing resource.

SELECT X, Y, (X-Y)^2 from ftable;

SELECT X, Y, Pcol_1 from ftable;

Just reference to (SELECT X, Y, (X-Y)^2 AS Pcol_1
the calculated result from remote_data_source)
in the remote side
Remote Query

Direction
of
The Future
Development


Move to OpenCL - WIP

• OpenCL support, instead of CUDA
• multiplatform support
• built-in JIT compiler

OpenCL Source
(CString)

clCreateProgramWithSource() ○ ○
cl_program

clCreateKernel() ○ ×
cl_kernel
○ ×
clEnqueueNDRangeKernel


Variable Length Data Support - WIP

• Data layout on chunk-buffer is revising, to accept variable-length data.
• Older format assumed fixed-number of items per chunk.
• Newer format assumes fixed-size chunk; consumed from head/tail

to consume
Direction
for fixed-length values X for fixed-length variable A

for index of variable-length value B
offset: 123
for fixed-length values Y

to consume
Direction
text: ‘hello world’
for contents of variable-lengths
for fixed-length values Z

Older chunk-buffer layout Newer chunk-buffer layout

Procedural Language Support

• This idea allows users to describe complicated logic
as procedural language to be executed on GPU.
• Expected usage: image processing, genome matching, ...

CREATE FUNCTION genome_similarity(text,text) RETURNS float AS
$$
varlena *genome1 = ARG1;
varlena *genome2 = ARG2;
:
<something complicated logic>
:
return similarity;
$$ LANGUAGE pg_strom;

SELECT id, label FROM genome_db
WHERE genome_similarity(data, ‘ATGCAGGT....’) > 0.9;


You getting involved

I’d like to know ...
• How PG-Strom run on
real-world dataset and workload
• How PG-Strom should get evolved
• Which region and problem will fit

All the co-development / co-evaluation
projects are always welcome!

Summary

• Characteristics of GPU & OpenCL
• Inflexible instructions but much higher parallelism
• Cheap and small power consumption per computing capability
• PG-Strom - towards most cost-effective database
• Utilization of GPU to off-load CPU jobs
• Automatic code generation and JIT compile
• Asynchronous execution
• Column-oriented data structure
• Upcoming development
• Move to OpenCL rather then CUDA
• Support for variable length values
• Support for procedural language
• Your involvement will lead future evolution!

Source

• Source code
• https://github.com/kaigai/pg_strom
• Wikipage
• http://wiki.postgresql.org/wiki/PGStrom
(need maintenance...)


Any Questions?


PG-Strom

More Related Content

What's hot (20)

Similar to PG-Strom (20)

More from Kohei KaiGai (20)

Recently uploaded (20)

PG-Strom