Brytlyt is GPU accelerated database that is based on PostgreSQL 9.4 and uses a Massively Parallel Processing (MPP) architecture to provide horizontal scale out for handling large amounts of data. Brytlyt is first released in 2016, now located at the milestone 2.0 (Dec.2018) and it has a long-term roadmap to version 5.0.
Multi-version Concurrency Control (MVCC)
Note that the Brytlyt database is derived from PostgreSQL. According to the lecture given by Richard Heyns, the CEO of Brytlyt, Brytlyt team does not change the logic in the concurrency control level. It applies Multi-version Concurrency Control for data consistency. For MVCC, not only the current status but also previous values of data are visible to the transaction, which provides transaction isolations. The primary advantage of MVCC overlocking is that the writing operation won't conflict with the reading operation on the same block of data. Thus, MVCC reduces the lock contention to achieve high throughput.
The traditional approach for running joins on CPU and is not well suited for the hundreds of thousands of cores in a GPU system. Since GPU’s have cores grouped in chunks, with each chunk running the same instructions, most GPU Databases have a tough time with join operations.
Brytlyt has approached the parallelism challenge by devising a patent-pending method that recursively separates rows containing a hit from rows that do not. It breaks the data into blocks and then distributes the blocks to the many cores used for searching.
For example, a dataset of 400,000 rows would be broken into blocks of 200 rows on a 2000-core GPU. Each GPU core then runs its own search on its own block of data in parallel with all the other cores, giving a huge boost in performance over the traditional CPU Database.
Empty blocks are discarded, and the process repeated with the remaining blocks. Then the whole process is done over and over until only the relevant blocks remain. This is an easily scalable process, and the importance of that cannot be overestimated. 10 billion rows could be distributed over 100 GPUs and achieve exactly the same cycle time as 1 billion rows on 10 GPUs.