PGCon2015 - final

PGCon 2015
The PostgreSQL Conference

KaiGai Kohei
Day Talks - Day 4 - Friday - 2015-06-19
Room DMS 1120
Start time 14:00
Duration 00:45
ID 801
Event type Lecture
Track Performance
Language used for presentation English


GPGPU meets PostgreSQL to accelerate analytic queries

The upcoming v9.5 supports custom-scan interface that allows extensions to provide alternative logic to scan/join relations. PG-Strom is an extension to off-load a part of CPU intensive workloads to GPU, built on top of the custom-scan interface. At this moment, it supports scan, join, aggregate and sorting to execute on GPU devices, and records x10 time faster response time in some usual queries.

Upcoming PostgreSQL v9.5 will support custom-scan interface that enables extensions to implement alternative logic to scan and/or join relations, then it shall be executed if these alternative logic is cheaper than built-in paths. PG-Strom is an extension built on top of the custom-scan interface, to process a part of CPU intensive SQL workloads on GPU processors. This feature works transparently from the application perspective; it internally generates native GPU binary and run the executable on GPU devices in asynchronous manner.

GPU has long-standing technology in HPC region, also different characteristics towards usual CPUs. It allocates much larger percentage of semiconductor chip for ALU logic, rather then cache or control units, therefore, it has much higher computing capability around TFLOPS grade on simple/massive numeric calculation, but not good at complicated logic. Recent GPUs mounts multi-hundreds to multi-thousands cores within a chip, thus much higher performance-cost ratio.

PG-Strom intermediates a world of SQL and a world of GPU. It makes great advantage for people in SQL world through the parallel query processing in low cost.

This session introduce the brief design of PG-Strom, background technology (custom-scan interface and GPU/OpenCL), current functionality and future development; focus on technology perspective. We expect audience are interested in data-analysis, OLAP or big-data.