Speed at Scale: Benchmarking GPU-Accelerated HeavyDB
From the very beginning, HEAVY.AI was built as a vertically integrated platform designed to allow fully interactive visual exploration of massive multi-billion row datasets, without needing to index, downsample, or pre-aggregate the data. This required the development of a built-in capability to zero-copy render large datasets using the GPU, as well as a web-based visual analytics frontend, HeavyImmerse, that would allow users to effortlessly slice-and-dice data with a few clicks and drags of their mouse.
However none of the blistering speeds the platform is known for would be possible without the performance of its core engine, the GPU-accelerated analytics database HeavyDB. Not only was HeavyDB built from the ground up to leverage the unique architecture of Nvidia GPUs, but we architected it to ensure it could exploit the full parallelism of modern hardware, both GPU and CPU. Here is a non-exhaustive list of key architectural innovations of the system:
While we often focus on the capabilities of the full vertical stack, over time our core engine has improved significantly both in terms of performance and capabilities. Hence we thought it would be interesting to test the raw performance of the system and compare against leading CPU data warehouses, using the widely used Star Schema Benchmark (SSB) and TPC-H tests. We think the results are impressive (with still further room for optimization), but we’ll let you judge for yourself.
