PG-Strom v2.0 Technical Brief (17-Apr-2018)

PG-Strom v2.0 Release
Technical Brief
(17-Apr-2018)
PG-Strom Development Team
<pgstrom@heterodb.com>

What is PG-Strom?
PG-Strom v2.0 Release Technical Brief (17-Apr-2018)2
GPU
[For I/O intensive workloads]
SSD-to-GPU Direct SQL Execution
In-memory Columnar Cache
SCAN+JOIN+GROUP BY
combined GPU kernel
[For Advanced Analytics workloads]
CPU+GPU hybrid parallel
PL/CUDA user defined function
GPU memory store (Gstore_fdw)
off-loading
✓ Designed as PostgreSQL extension
✓ Transparent SQL acceleration
✓ Cooperation with fully transactional
database management system
✓ Various comprehensive tools and
applications for PostgreSQL
PG-Strom: an extension module to accelerate analytic SQL workloads using GPU.
[GPU’s characteristics]
✓ Several thousands of processor cores per device
✓ Nearly terabytes per second memory bandwidth
✓ Much higher cost performance ratio
PG-Strom is an open source extension module for PostgreSQL (v9.6 or later) to accelerate analytic queries, has been developed and
incrementally improved since 2012. It provides PostgreSQL alternative query execution plan for SCAN, JOIN and GROUP BY workloads.
Once optimizer chose the custom plans which use GPU for SQL execution, PG-Strom constructs relative GPU code on the fly. It means
PostgreSQL uses GPU aware execution engine only when it makes sense, so here is no downside for transactional workloads.
PL/CUDA is a derivational feature that allows manual optimization with user defined function (UDF) by CUDA C; which shall be executed on
GPU device. The PG-Strom v2.0 added many features to enhance these basis for more performance and wider use scenarios.

PG-Strom v2.0 features highlight
▌Storage Enhancement
 SSD-to-GPU Direct SQL Execution
 In-memory Columnar Cache
 GPU memory store (gstore_fdw)
▌Advanced SQL Infrastructure
 PostgreSQL v9.6/v10 support – CPU+GPU Hybrid Parallel
 SCAN+JOIN+GROUP BY combined GPU kernel
 Utilization of demand paging of GPU device memory
▌Miscellaneous
 PL/CUDA related enhancement
 New data type support
 Documentation and Packaging

Storage enhancement for each layer
Storage Layer
GPU device memory
Band: ~900GB/s
Capacity: ~32GB
Host memory
Band: ~128GB/s
Capacity: ~1.5TB
NVMe-SSD
Band: ~10GB/s
Capacity: 10TB and more
Hot Storage
Cold Storage
Related PG-Strom Feature
GPU memory store
(gstore_fdw)
In-memory columnar cache
SSD-to-GPU Direct SQL
Execution
In general, key of performance is not only number of cores and its clock, but data throughput to be loaded for the processors also.
Individual storage layer has its own characteristics, thus PG-Strom provides several options to optimize the supply of data.
SSD-to-GPU Direct SQL Execution is unique and characteristic feature of PG-Strom. It directly loads the data blocks of PostgreSQL to GPU,
and runs SQL workloads to reduce amount of data to be processed by CPU prior to the arrival. Its data transfer bypasses operating system
software stacks, thus allows to pull out nearly wired performance of the hardware. In-memory columnar cache allows to keep data blocks
in the optimal format for GPU to compute and transfer over the PCIe bus. GPU memory store (gstore_fdw) allows preload on the GPU
device memory using standard SQL statement. It is an ideal data location for PL/CUDA function because it does not need to carry the data
set for each invocation, and no size limitation of 1GB which is maximum length of the varlena.

SSD-to-GPU Direct SQL Execution (1/3)
Pre-processing of SQL workloads to drop unnecessary rows prior to the data loading to CPU
GPU code generation
PG-Strom automatically
generate CUDA code. It
looks like a transparent
acceleration from the
standpoint of users.
SQL execution on GPU
It loads data blocks of
PostgreSQL to GPU using
peer-to-peer DMA, then
drops unnecessary data
with parallel SQL execution
by GPU.
SQL execution on CPU
CPU runs pre-processed
data set; that is much
smaller than the original.
Eventually, it looks like
GPU accelerated I/O also.
PCIe Bus
NVMe SSD GPU
SSD-to-GPU P2P DMA
(NVMe-Strom driver)
WHERE-clause
JOIN
GROUP BY
Large PostgreSQL
Tables
PostgreSQL
Data Blocks
SSD-to-GPU Direct SQL
Once data blocks are loaded to GPU,
we can process SQL workloads using
thousands cores of GPU. It reduces
the amount of data to be loaded and
executed by CPU; looks like I/O
performance acceleration.
Traditional data flow
Even if unnecessary records, only CPU
can determine whether these are
necessary or not. So, we have to move
any records including junks.
SELECT cat, count(*), avg(X)
FROM t0 JOIN t1 ON t0.id = t1.id
WHERE YMD >= 20120701
GROUP BY cat;
SQL optimization stage
SQL execution stage
SQL-to-GPU
Program
Generator
GPU binary
Just-in-time
Compile

0
1000
2000
3000
4000
5000
6000
7000
8000
Q1-1 Q1-2 Q1-3 Q2-1 Q2-2 Q2-3 Q3-1 Q3-2 Q3-3 Q3-4 Q4-1 Q4-2 Q4-3
QueryExecutionThroughput[MB/s]
Star Schema Benchmark results on NVMe-SSDx3 with md-raid0
PostgreSQL v10.3 PG-Strom v2.0
This result shows query execution throughput on 13 different queries of the Star Schema Benchmark. Host system mounts just 192GB
memory but database size to be scanned is 353GB. So, it is quite i/o intensive workload.
PG-Strom with SSD-to-GPU Direct SQL Execution shows 7.2-7.7GB/s throughput which is more than x3.5 faster than the vanilla PostgreSQL
for large scale batch processing. The throughput is calculated by (database size) / (query response time). So, average query response time of
PG-Strom is later half of the 40s, and PostgreSQL is between 200s to 300s.
Benchmark environment:
Server: Supermicro 1019GP-TT, CPU: Intel Xeon Gold 6126T (2.6GHz, 12C), RAM: 192GB, GPU: NVIDIA Tesla P40 (3840C, 24GB),
SSD: Intel DC P4600 (HHHL, 2.0TB) x3, OS: CentOS 7.4, SW: CUDA 9.1, PostgreSQL v10.3, PG-Strom v2.0

Tesla GPU
NVIDIA
CUDA
Toolkit
Filesystem
(ext4, xfs)
nvme driver (inbox)
nvme_strom
kernel
module
NVMe SSD drives
commodity x86_64 hardware
NVIDIA GPUDirect RDMA
NVIDIA
kernel
driver
PostgreSQL
pg_strom
extension
read(2) ioctl(2)
Hardware
Layer
Operating
System
Software
Layer
Database
Software
Layer
Application Software
SQL Interface
I/O path based on
normal filesystem
I/O path based on
SSD-to-GPU Direct SQL Execution
SSD-to-GPU Direct SQL Execution is
a technology built on top of NVIDIA
GPUDirect RDMA which allows P2P
DMA between GPU and 3rd party PCIe
devices.
The nvme_strom kernel module
intermediates P2P DMA from NVMe-
SSD to Tesla GPU [*1].
Once GPU accelerated query execution
plan gets chosen by the query optimizer,
then PG-Strom calls ioctl(2) to deliver
the command of SSD-to-GPU Direct SQL
Execution to nvme_strom in the kernel
space.
This driver has small interaction with
filesystem to convert file descriptor +
file offset to block numbers on the
device. So, only limited filesystems
(Ext4, XFS) are now supported.
You can also use striping of NVMe-SSD
using md-raid0 for more throughput,
however, it is a feature of commercial
subscription. Please contact HeteroDB,
if you need multi-SSDs grade
performance.
[*1] NVIDIA GPUDirect RDMA is available on
only Tesla or Quadro, not GeForce.

In-memory Columnar Cache
PostgreSQL
heap buffer
(row-format)
In-memory
columnar
cache
(column-format)
GPU kernel
for column
format
GPU kernel
for row
format
GPU kernel
for column
format
GPU kernel
for column
format
GPU kernel
for row
format
Query Execution on GPU device
sequential scan
asynchronous columnar cache builders (background workers)
transactional workloads
UPDATE
cache
invalidation
on write

(Background) Why columnar-format is preferable for GPU
▌Row-format – random memory access
▌Column-format – coalesced memory access
32bit
Memory transaction width: 256bit
32bit 32bit32bit 32bit 32bit
32bit 32bit 32bit 32bit 32bit 32bit 32bit 32bit
Memory transaction width: 256bit
32bits x 8 = 256bits are valid
in the 256bits memory transaction
(usage ratio: 100.0%)
Only 32bits x 1 = 32bits are valid
in the 256bits memory transaction
(usage ratio: 12.5%)
GPU cores
GPU cores
Memory access pattern affects to GPU’s performance so much because of its memory sub-system architecture. GPU has relatively larger
memory transaction width. If multiple co-operating cores simultaneously read continuous items in array, one memory transaction can load
eight 32bit values (if width is 256bit) at once. It fully depends on the data format, and one of the most significant optimization factor.
Row-format tends to put values of a particular column to be scanned on random place, not continuous, thus it is hard to pull out maximum
performance of GPU. Column-format represents a table like as a set of simple array. So, when column-X is referenced in scan, all the values
are located very closely thus tend to fit the coalesced memory access pattern.

GPU computing world
GPU memory store (Gstore_fdw)
Storage
SQL world
GPU device memory
Foreign Table
(gstore_fdw)
INSERT
UPDATE
DELETE
SELECT
Reference
by Zero-copy
✓ Data format conversion
✓ Data compression
✓ Transaction management
PL/CUDA
user defined
functions
Gstore_fdw provides a set of interface to read/write GPU device memory using standard SQL. You can load bulk data into GPU’s device
memory using INSERT command, and so on. Because all the operations are handled inside PostgreSQL database system, data keep its binary
form (so, no needs to dump as CSV file once and parse by Python script again). SQL is one of the most flexible tool for data management, so
you can load arbitrary data-set from the master table, and apply pre-processing required by machine-learning algorithm on the fly.
One other significant feature is data-collaboration with external programs like Python scripts. GPU device memory can be shared with
external program once identifier of the acquired memory region (just 64bytes token) is exported. It enables to use PostgreSQL as a powerful
data management infrastructure for machine-learning usage. Even though Gstore_fdw now supports only ‘pgstrom’ internal format, we will
support other internal format in the next or future version.
IPC Handle
User written
Scripts
Machine Learning
Framework
IPC Handle

PostgreSQL v9.6/v10 support – CPU+GPU Hybrid Parallel
SeqScan
Part-Agg
HashJoin
SeqScan
Part-Agg
HashJoin
SeqScan
Part-Agg
HashJoin
Gather
Final-Agg
GpuScan
GpuPreAgg
GpuJoin
Gather
Final-Agg
GpuScan
GpuPreAgg
GpuJoin
GpuScan
GpuPreAgg
GpuJoin
result result
CPU Parallel Execution CPU + GPU Hybrid Parallel Execution
CPU parallel execution per process granularity CPU parallel execution per process granularity
PostgreSQL’s worker
process individually
uses GPU for finer
grained granularity.
 Multi-process pulls up GPU usage more efficiently
One epoch-making feature at PostgreSQL v9.6 was parallel query execution based on concurrent multi-processes. It also extended the
custom-scan interface for extensions to support parallel-query. PG-Strom v2.0 was re-designed according to the new interface set, then
it enables GPU aware custom-plan to run on the background worker process.
Heuristically, capacity of single CPU thread is not sufficient to supply enough amount of data stream to GPU. It is usually much narrower
than GPU’s computing capability. So, we can expect CPU parallel execution assists to pull up GPU usage with much higher data supply ratio.

SCAN + JOIN + GROUP BY combined GPU kernel
CustomScan interface allows extension to inject alternative implementation of query execution plan. If estimated cost is more reasonable
than built-in implementation, query optimizer of PostgreSQL chooses the alternative plan. PG-Strom provides its custom logic for SCAN,
JOIN and GROUP BY workloads, with GPU acceleration.
Usually, JOIN consumes the result of SCAN then generates records as its output. The output of JOIN often performs as input of GROUP BY,
then it generates aggregation.
In case when GpuPreAgg, GpuJoin and GpuScan are continuously executed, we have an opportunity of further optimization by reduction of
the data transfer over PCIe bus. The result of GpuScan can perform as GpuJoin’s input, and the result of GpuJoin can also perform as
GpuPreAgg’s input, if all of them shall be continuously executed. PG-Strom tries to re-use the result buffer of the previous step as input
buffer of the next step. It allows to eliminate the data ping-pong over the PCIe-bus. Once SCAN + JOIN + GROUP BY combined GPU kernel
gets ready, it can run with the most efficient GPU kernel because it does not need data exchange over the query execution plan.
GpuScan
kernel
GpuJoin
kernel
GpuPreAgg
kernel
Agg
(PostgreSQL)
GPU
CPU
Storage
Results
SCAN + JOIN + GROUP BY Combined Kernel
Data
Blocks
Host
Buffer
Host
Buffer
elimination of
data ping-pong

Utilization of demand paging of GPU device memory
Buffer size estimation is not an easy job for SQL workloads. In general, we cannot know exact number of result rows unless query is not
actually executed, even though table statistics informs us rough estimation through the query planning.
GPU kernel needs result buffer for each operation, and it has to be acquired prior to its execution. It has been a problematic trade-off
because a tight configuration with small margin often leads lack of the result buffer, on the other hands, buffer allocation with large margin
makes unignorable dead space on the GPU device memory.
The recent GPU (Pascal / Volta) supports demand paging of GPU device memory. It assigns physical page frame on demand, thus, unused
region consumes no physical device memory. It means that large margin configuration consumes no dead physical device memory.
It also allows to simplify the code to estimate the size of result buffer and to retry GPU kernel invocation with larger result buffer. These
logics are very complicated and had many potential bugs around the error pass.
PG-Strom v2.0 fully utilized the demand paging of GPU device, thus we could eliminate the problematic code. It also contributed the
stability of the software.
Hash Table t2
Hash Table t1
Data chunk of
table t0
t0
GpuJoin
t1
t2
result buffer
Kepler / Maxwel
 Dead space (!)
Pascal / Volta
 No physical page frames
are not assigned, so harmless

PL/CUDA related enhancement
All In-database Analytics Scan
Pre-Process
Analytics
Post-ProcessCREATE FUNCTION
my_logic( reggstore, text )
RETURNS matrix
AS $$
$$ LANGUAGE ‘plcuda’;
Custom CUDA C code block
(runs on GPU device) ✓ manually optimized analytics &
machine-learning algorithms
✓ utilization of a few thousands
processor cores
✓ ultra high memory bandwidth
If “reggstore” type is supplied as argument of PL/CUDA function, user defined part of this PL/CUDA function receives this argument as
pointer to the preserved GPU device memory for gstore_fdw. It allows to reference the data preliminary loaded onto the Gstore_fdw.
The #plcuda_include directive is newly supported to include the code which is returned from the specifies function. You can switch the
code to be compiled according to the arguments, not to create multiple but similar variations.
CREATE FUNCTION my_distance(reggstore, text)
RETURNS text
AS $$
if ($2 = “manhattan”)
return “#define dist(X,Y) abs(X-y)”;
if ($2 = “euclid”)
return “#define dist(X,Y) ((X-y)^2)”
$$ ...
#plcuda_include my_distance
② Inclusion of other function’s result
as a part of CUDA C code.
GPU memory
store
(gstore_fdw)
① reggstore argument as reference to gstore_fdw structure

New data type support
▌Numeric
 float2
▌Network address
 macaddr, inet, cidr
▌Range data types
 int4range, int8range, tsrange, tstzrange, daterange
▌Miscellaneous
 uuid, reggstore
The “float2” data type is implemented by PG-Strom, not a standard built-in data type. It represents half-precision floating point values.
People in machine-learning area often use FP16 for more short representation of matrix; less device memory consumption and higher
computing throughput.

Documentation and Packaging
Documentation was totally rewritten with markdown that is much easier for timely update than raw HTML based one.
It is now published at http://heterodb.github.io/pg-strom/
RPM packages are also available for RHEL7.x / CentOS 7.x. PG-Strom and related software are available on the HeteroDB Software
Distribution Center (SWDC) at https://heterodb.github.io/swdc/
You can use the SWDC as yum repository source.
PG-Strom official documentation HeteroDB Software Distribution Center

PostgreSQL v11 support (1/2) – Parallel Append & Multi-GPUs
 PG10 Restriction: GPUs not close to SSD must be idle during the scan across partition table.
 PG11 allows to scan partitioned children in parallel. It also makes multiple-GPUs active simultaneously.
Parallel
Append
t1 t2 t4t3
GPU1 GPU2
BgWorker BgWorker BgWorker BgWorker
Background worker process that
contains GPU-aware custom-plans
Choose the GPU1 which shares
the same PCIe root complex of
the NVMe-SSD to be scanned.
auto-tuning
based on PCIe
topology
NVIDIA GPUDirect RDMA, is a basis technology of SSD-to-GPU Direct SQL Execution, requires GPU and SSD should share the PCIe root
complex, thus P2P DMA route should not traverse QPI link. It leads a restriction on multi-GPUs configuration with partitioned table.
When background worker scans partitioned child tables across multiple SSDs, GPU-aware custom plan needs to choose its neighbor GPU to
avoid QPI traversal. In other words, the target table for scan determines the GPU to be attached on the background workers.
In PG10, partitioned child tables are sequentially picked up to scan, so we cannot activate more than one GPUs simultaneously because the
secondary GPU will cause QPI traverse on usual hardware configuration (1:CPU – 1:GPU + n:SSDs).
PG11 supports parallel scan across partitioned child tables. It allows individual background worker activate its neighbor GPU for the tables
they are scanning. It enables to utilize multiple GPUs under SSD-to-GPU Direct SQL Execution for larger data processing.

PostgreSQL v11 support (2/2) – In-box distributed query execution
PCIe I/O Expansion Box
Host System
(x86_64 server)
NVMe SSD
PostgreSQL Tables
PostgreSQL
Data Blocks
Internal
PCIe Switch
SSD-to-GPU P2P DMA
(Large data size)
GPU
WHERE-clause
JOIN
GROUP BY
PCIe over
Ethernet
Pre-processed
small data
Several GB/s
SQL execution
per Box
Several GB/s
SQL execution
per Box
Several GB/s
SQL execution
per Box
Add performance and capacity
NIC / HBA
Performs as just a simple single node
configuration from the standpoint of
applications and administrators
Multi-GPUs capability will expand the opportunity of big data processing for more bigger data set, probably, close to 100TB.
Multiple vendors provide PCIe I/O expansion box solution that allows to install PCIe devices on the physically separated box which is
connected to the host system using fast network. In addition, some of the solution have internal PCIe switch that can route P2P DMA packet
inside of the I/O box. It means PG-Strom handles SSD-to-GPU Direct SQL Execution on the I/O box with little interaction to the host system,
and runs the partial workload on the partitioned table per box in parallel once multi-GPUs capability get supported at PG11.
From the standpoint of applications and administrators, it is just a simple single node configuration even though many GPUs and SSDs are
installed, thus, no need to pay attention for distributed transaction. It makes application design and daily maintenance so simplified.

Other significant features
▌cuPy data format support of Gstore_fdw
▌BRIN index support
▌Basic PostGIS support
▌NVMe over Fabric support
▌GPU device function that can return varlena datum
▌Semi- / Anti- Join support
▌MVCC visibility checks on the device side
▌Compression support of in-memory columnar cache
See 003: Development Roadmap for more details.

PG-Strom v2.0 Technical Brief (17-Apr-2018)

More Related Content

What's hot (20)

Similar to PG-Strom v2.0 Technical Brief (17-Apr-2018) (20)

More from Kohei KaiGai (20)

Recently uploaded (20)

PG-Strom v2.0 Technical Brief (17-Apr-2018)