Pattern-Oriented Compiler Optimization In Datacenters

We are moving into the era of multi/many-core and cloud computing. On one side, more and more on-chip parallelism is introduced, how to utilize the tremendous computing power become a key problem and a main task of modern compiler systems. On the other side, multiple applications would co-run simultaneously, which breaks the premise "the real execution environments of applications match with what we suppose at compilation time" and brings new challenge to compilation method.This paper implements "Extendable Pattern-Oriented Optimization Directives (EPOD)" on LLVM platform and proposes a new dynamic peer-aware compilation methodology for datacenters. This paper makes the following contributions:1) We implemented EPOD on LLVM platform, including making and recognizing semantic pattern, defining specialized optimization strategy for the semantic pattern;2) We added a new semantic pattern: Matrix Multiplication to EPOD;3) We proposed a new dynamic peer-aware compilation methodology, we first modeled the relationship between an application’s optimal optimization strategy and the pressure on shared resources, and developed schemes to dynamically adjust an application’s optimization strategy according to the execution environments. At last, we evaluated our optimization strategy for matrix multiplication and peer-aware compilation methodology. Evaluations showed that our proposed EPOD-based matrix multiplication could achieve43X speedup over commercial compilers for both general and special matrix multiplication. For peer-aware compilation methodology, we also achieved25.3%speedup over static compilation methodology.