| layout | title | permalink |
|---|---|---|
default |
Domain Specific Run Time Optimization for Software Data Planes |
/ |
Morpheus is a system working alongside static compilers that continuously optimizes the targeted networking code. It introduces a number of new techniques, from static code analysis to adaptive code instrumentation, together with a toolbox of domain specific optimizations used to manipulate the code on-the-fly depending on runtime traffic patterns and control plane configurations.
Morpheus compilation pipeline.
The Morpheus core exploits the LLVM compiler toolchain (v10.0.1) for code manipulation and run-time code generation. It works at the intermediate representation (IR) level as it allows to reason about the running code using a relatively high-level language framework without compromising on code generation time.
The Morpheus compilation pipeline is characterized of three main steps:
- Code Analysis: In the first pass, Morpheus uses static code analysis to identify all map access sites in the code, understand whether a particular access is a read or a write operation, and reason about the way the result is used later in the code. In particular, signature-based call site analysis is used to track map lookup and update calls, and then a combination of memory dependency analysis and alias analysis is performed to match map lookups to map updates.
- Instrumentation: In the second pass, Morpheus profiles the dynamics of the input traffic by generating heatmaps of the maps’ access patterns, so that the collected statistics can then be used to drive the subsequent optimization passes.
- Optimization Passes: The third step of the compilation pipeline is where all online code transformations are applied (e.g., JIT compilation, Dead code elimination, Constant Propagation, Guard Elision).
For more information, please read the Morpheus's paper and the extended abstract accepted to ASPLOS '22.
We have prepared Morpheus for artifact evaluation at ASPLOS'22. Here is the link to all artifacts in the paper.