PolarDB:Enable GPU-accelerated computing - PolarDB

GanosBase automatically enables GPU-accelerated computing when it detects a GPU in the environment. No configuration is required. You can also control GPU acceleration at the session level to optimize performance for specific workloads.

How it works

GPU-based parallel acceleration in GanosBase operates at the object level. GanosBase converts a field object into a model suitable for parallel computing, then distributes the computation across multiple GPU cores. This gives GPUs a significant advantage over CPUs for compute-intensive and embarrassingly parallel workloads.

Prerequisites

Before you begin, ensure that you have:

A GPU installed and accessible in your PolarDB cluster environment

Check GPU availability

Create the ganos_raster extension:
```
create extension ganos_raster cascade;
```

Run st_checkgpu() to check whether a GPU is available:

GPU-accelerated computing requires a GPU in the environment.

select st_checkgpu();

If a GPU is detected, GanosBase returns the GPU properties:

rasterdb=# select st_checkgpu();
                             st_checkgpu
------------------------------------------------------------------------------------------------------------------------------
 [GPU(0) prop]multiProcessorCount=20; sharedMemPerBlock=49152; totalGlobalMem=-608239616; maxThreadsPerBlock=1024; maxThreadsPerMultiProcessor=2048; cudaThreadGetLimit=1024 .
(1 row)

If no GPU is detected, GanosBase returns an error message:

rasterdb=# select st_checkgpu();
                           st_checkgpu
------------------------------------------------------------------------------------------------------------------------------------------
There is not gpu device on current environment, cuda_errorcode=35, errormsg=CUDA driver version is insufficient for CUDA runtime version.
(1 row)

Control GPU acceleration at the session level

By default, GanosBase enables GPU acceleration in any environment where a GPU is available.

Disable GPU acceleration (switch to CPU mode)

For highly concurrent workloads, a single GPU may become a bottleneck. Disable GPU acceleration at the session level to route those queries to CPU:

set ganos.raster.use_cuda=off;

Verify the setting:

show ganos.raster.use_cuda;

Expected output:

 ganos.raster.use_cuda
-----------------------
 off
(1 row)

Re-enable GPU acceleration

set ganos.raster.use_cuda=on;

Verify the setting:

show ganos.raster.use_cuda;

Expected output:

 ganos.raster.use_cuda
-----------------------
 on
(1 row)

Limitations

GPU-accelerated computing is currently supported only in the GanosBase Raster module. Support for the GanosBase Trajectory and GanosBase Geometry modules is planned for a future release.