PgConf.Russia 2016

Postgresql's buffer manager has parts where it's showing its age. We'll discuss how it currently works, what problems there are, and what attempts are in progress to rectify its weaknesses.

• Lookups in the buffer cache are expensive
• The buffer mapping table is organized as a hash table, which makes efficient implementations of prefetching, write coalescing, dropping of cache contents hard
• Relation extension scales badly
• Cache replacement is inefficient
• Cache replacement replaces the wrong buffers

When new versions of Postgres are released most of the attention is focused on new features. Inevitably a release note claiming speed improvements seems relatively mundane and doesn't provide the compelling argument for upgrading. However the reality is that these speed improvements represent pain points that have been identified and solved.

Reviewing the changes to the sort code in Postgres over the last 10 years clearly shows the kinds of problems users have run into. As usage patterns changed over years, databases scaled up, and hardware changed new problems arose and drove further development to solve them.

Upcoming changes in 9.5 and 9.6 will dramatically change the experience further. Making sorting UTF8 and other encodings less of a problem and handling scaling to larger machines with many processors and memory cache more effectively.

Currently, PostgreSQL uses the interpreter to execute SQL-queries. This yields an overhead caused by indirect calls to handler functions and runtime checks, which could be avoided if the query were compiled into the native code "on-the-fly" (i.e. JIT-compiled): at a run time the specific table structure is known as well as data types used in the query. This is especially important for complex queries, which performance is CPU-bound. At the moment there are two major projects that implement JIT-compilation in PostgreSQL: a commercial database Vitesse DB and an open-source project PGStorm. The former uses LLVM JIT to achieve up to 8x speedup on selected TPC-H benchmarks, while the latter JIT-compiles the query using CUDA and executes it on GPU, which allows to speed up execution of specific query types by an order.

Our work is dedicated to adding support for SQL query JIT-compilation to PostgreSQL using LLVM compiler infrastructure. In the presentation we'll discuss how JIT-compilation can be used to speed up various stages of query execution in PostgreSQL, and the specifics of translating an SQL query into LLVM bitcode to achieve good performing native code. Also we'll present preliminary results for our JIT-compiler on TPC-H benchmark.