# 1. Background ## VectorWise - PreCompiled Primitives - 预编译 primitives - primitive：在 typed data 上执行简单的操作 - 操作简单意味着更容易向量化 - DBMS 在执行 query plan 时运行时调用 primitives（通过函数调用） - 一次处理多个 tuples 来均摊函数调用的开销 - primitive 的输出：满足 predicate 的 tuple offsets（类似 bitmap） -  ## Hyper - Holistic Query Compilation - 使用 LLVM 将查询编译成 native code - Push-based query processing - 尽可能地将 tuple 保持在 CPU 寄存器中 -------- # 2. Vectorization vs. Compilation - [Everything you always wanted to know about compiled and vectorized queries but were afraid to ask](https://dl.acm.org/doi/abs/10.14778/3275366.3284966) ## Implementations - 方式一：**Tectorwise** - 将操作分解为 pre-compiled primitives - 每一步必须 materialize primitives 的输出，再传递给下一个 primitive - 方式二：**Typer** - Push-based processing mode，加上 JIT 编译 - 不物化中间结果，Process a single tuple up entire pipeline ## TPC-H Workload - [TPC-H Analyzed: Hidden Messages and Lessons Learned from an Influential Benchmark](https://homepages.cwi.nl/~boncz/snb-challenge/chokepoints-tpctc.pdf) - [cmu-db/benchbase](https://github.com/cmu-db/benchbase/tree/main/src/main/java/com/oltpbenchmark/benchmarks/tpch/procedures) - Java TPC-H queries ## Main Findings - 两种模型都很高效，达到差不多相同的性能 - 比行存 DBMSs 快了 100 倍 - Data-centric 更适合 calculation-heavy 的查询，cache misses 更少 - Vectorization 在 hiding cache miss latencies 方面稍微好一些 - 一次处理多个 tuples ## SIMD Preformance - AVX-512 有专门的 bitmask registers，适合 selective operations ## Auto-Vectorization - 评估编译器在自动向量化 Vectorwise primitives 时的表现 - 目标：GCC v7.2，Clang v5.0，ICC v18 ## Vectorization VS. Compilation -  ------------ # Parting Thoughts > No major performance difference between the Vectorwise and HyPer approaches for all queries. - 大多数 OLAP 数据库都选择了向量化，更容易实现、Debug