PGStrom for Postgres

• PGStrom
  • a tool to execute a special type of SQL query on the GPU.
  • 50x compared to standard Postgres
    • because they use the GPU
• PGStrom is faster than Postgres,
  • but this is not because of the GPU.
• a multi-core CPU can process this queries as fast as the GPU.
• For the test-case described on the PGStrom wiki page the CPU is even a bit faster.

Introduction

• colleagues mailed me the link to PGStrom,
  • they knew that this is in my domain
  • 10x speedups by using a GPU in my field of work usually makes me skeptical.
• PGStrom executes statements in the form of

SELECT * FROM table WHERE <complex formula> 

…on the GPU.

• This is not a bad idea at all.
• If you take a look at my last blog post, you can see that we have easily 5 checks out of 6.
• The algorithm can be processed in a SIMD fashion, streamed without overhead and we do not need to transfer data more than once (a.k.a. “fits into GPU memory”).
• However, the most important check is missing, and that’s the part where I am skeptical: I am not sure if the transfer to the GPU is faster than the computation on the CPU.

• not appropriate to compare Postgres against an optimized implementation on the GPU.
• Pg processes 10 Mio records in 7 seconds in their experiment.
  • 10 Mio sounds like a huge number, but actually the example statement processes only two columns with floating point numbers.
  • Assuming these numbers are 32 bit-float types, the CPU processes only 76 MB – in 7 seconds!
  • Remember that one CPU thread is capable of transferring around 4 GB of data per second!

Experiment and Implementation

• my opinion
  • a fair comparison of GPU vs CPU is to
  • hard-code this query in C and in CUDA and
  • compare the run-time.
• Because of the detailed description of the PGStrom test setup on the wiki-page (better than in most research papers…),
  • we can easily reproduce their tests. This SQL statement

SELECT COUNT(*) FROM t1 WHERE sqrt((x-25.6)^2 + (y-12.8)^2) < 15;