April 2021 Summaries

The KISS principle suggests using a simple solution when it works, but many people mistakenly believe that this means only using single-node DBMSs like MySQL, PostgreSQL, or SQL Server. However, scale-out databases, such as SingleStore, can actually simplify things by providing a scalable and performant solution for medium-to-large data sets, enabling digital transformation initiatives and unlocking the value of data. Scale-out databases are designed to be run across machines, which allows for vertical scaling up to larger machines or horizontal scaling out to multiple nodes, making them more suitable for applications with large data sets that require fast response times and high concurrency. By using a scale-out database, developers can avoid complex solutions like pre-calculated aggregate tables and materialized views, and instead leverage the power of sharding and reference tables to simplify their design. Additionally, modern SQL-based scale-out databases are often compatible with standard SQL and many existing languages and tools, making it easy for developers to adopt them and transfer their existing skills to a new solution.

Linus Torvalds' sentiment on AVX512 is that he hopes it "dies a painful death" due to his dislike of FP benchmarks, but this view is disputed by the author who has successfully migrated SingleStore's SIMD kernels from AVX2 to AVX512/VBMI. The author found that AVX512 provides more powerful instructions and predicated instructions, making it easier to use and reducing development time for developers. The new multishift instruction allows for arbitrary byte-level permutes in a single operation, making the code easier to read and less prone to errors. Additionally, the predicated instructions enable seamless conditional execution, allowing for efficient conversion between bit vectors and byte vectors. Overall, the author concludes that AVX512/VBMI is a more user-friendly SIMD instruction set, providing significant performance benefits and developer time savings.