19 KiB
| layout | title | category | tags | order | ||
|---|---|---|---|---|---|---|
| developer-doc | Runtime Features | runtime |
|
4 |
Runtime Features
This document contains a detailed specification of Enso's runtime. It includes a description of the technologies on which it is built, as well as the features and functionality that it is required to support. In addition, the document aims to explain why this design, rather than one of the many alternatives available to the team.
When we refer to the Enso 'runtime' in this document, we are referring to the combination of the language communication protocol, typechecker, optimiser, and interpreter. Though the interpreter itself has its own runtime, it is these components that make up Enso's runtime.
The runtime is built on top of GraalVM, a universal virtual machine on which you can run any language with an appropriate interpreter. In basing Enso's runtime on GraalVM, we not only have access to a comprehensive toolkit for building high-performance language interpreters, but also to the ecosystems of all the other languages (e.g. C++, Python, R) that can run on top of it. GraalVM also brings some additional important tooling, such as the JVM ecosystem's performance monitoring, analysis, and debugging toolsets.
The runtime described below is a complex beast, so this document is broken up into a number of sections. These aim to provide an architectural overview, and then describe the design of each component in detail.
- Architectural Overview
- Choosing GraalVM
- The Runtime Components
- Cross-Cutting Concerns
- The Initial Version of the Runtime
- Development Considerations
Architectural Overview
The Enso runtime is just one of the many components of the Enso ecosystem. This section provides an overview of how it fits into the broader ecosystem, with a particular focus on how it enables workflows for Enso Studio, the Enso CLI, and Language Server integration. In addition, this section also explores the architecture of the runtime itself, breaking down the opaque 'runtime' label into the
The Broader Enso Ecosystem
While the runtime is arguably the core part of Enso, for the language would not be able to exist without it, the language's success is just as dependent on the surrounding ecosystem.
TBC...
It is worth providing a brief explanation of each of the components to aid in understanding how the runtime fits into the ecosystem.
- Enso Studio GUI: This is the interface with which most of Enso's users will interact. It handles the drawing of and interaction with the Enso graph for users, as well as the searcher and other user-facing functionality. It also provides a text editor.
- Project Manager: This allows for management of one or more Enso projects, and is primarily responsible for file-system-agnostic interaction with the project structure, and spawning of the Enso runtime.
- GUI Backend: The GUI backend is instantiated for each project, and is
responsible for all of the user-facing logic that goes into interaction with
the Enso runtime.
- Graph State Manager: This component handles management of the state required to draw the graph in the GUI.
- Double Representation Manager: This component handles the encoding and decoding of the Enso program to and from the intermediate representation.
- Undo/Redo Manager: This component handles undo and redo for the graph, a somewhat novel operation as it does not not always have a 1:1 correspondence with textual editing.
- CLI: This provides a command-line (specifically a terminal) interface to the Enso runtime. This allows both for the CLI invocation of Enso, as well as an interactive REPL. This communicates with the runtime itself via the language server protocol.
- Enso Runtime: This is what is described in this document, and is responsible for the execution of Enso programs. It handles the typechecking, optimisation, and interpretation of Enso code, as well as the provision of interfaces to foreign languages.
The Runtime's Architecture
In order to better appreciate how the components specified below interact, it is important to have an understanding of the high-level architecture of the runtime itself. The design in this document pertains only to the 'Enso Runtime' component of the diagram above, and hence makes no mention of the others.
In the diagram below, the direction of arrows is used to represent the 'flow' of information between the various components.
TBC...
Choosing GraalVM
Building the runtime on top of GraalVM was of course not the only choice that could've been made, but it was overwhelmingly the most sensible option out of those considered.
At the time the runtime was designed, there were three main options that were being considered.
- LLVM: A battle-tested and comprehensive toolchain for the creation of language compilers, LLVM includes facilities for compilation, optimisation, JIT, and linking.
- GHC: The Glasgow Haskell Compiler is a sophisticated compiler and runtime for Haskell that provides a language-agnostic set of internal representations that could be leveraged to compile and/or interpret other functional languages.
- JVM: The JVM is a high-performance virtual machine that includes sophisticated garbage collection, profiling tools, and a JIT compiler.
- GraalVM: A universal virtual machine and language development toolkit, GraalVM provides a framework for building language interpreters, as well as a JIT compiler. Most importantly, it provides tools for seamless interoperability between languages that can run on Graal, which include Python and R.
The decision to build Enso's runtime using GraalVM was primarily motivated by business concerns, but these concerns did not override the technical as well. Addressing them one by one provides a comprehensive picture of why the decision was made.
Overall, it is clear that GraalVM is an optimal choice for Enso at this stage of the language's development. While the other potential targets do have their upsides (e.g. the JVM's sophisticated garbage collection machinery), they all had at least one 'fatal flaw' for Enso's use case.
Speed of Development
A language runtime is a complex beast, so any solution that could remove some of the implementation burden would be beneficial to Enso as a product.
Where LLVM provides comprehensive tools for compiling languages, it provides no actual runtime. This would require significant implementation effort, requiring the implementation of facilities for concurrency, as well as garbage collection, neither of which are simple tasks.
GHC, on the other hand, provides a comprehensive runtime system that includes both a garbage collector and sophisticated concurrency system. However, while it does provide language-agnostic intermediate representations, these are tied to Haskell from a development perspective. Unlike LLVM, GraalVM, or even the JVM, if GHC Haskell requires a change to these representations, that change will be made.
With many languages already targeting the JVM it also seemed like an attractive option. The stable bytecode target would be useful, but other languages have proven the challenges of generating sensible bytecode to provide good language performance.
GraalVM manages to provide excellent performance with a sensible, high-level interface, thereby enabling rapid development of a performant runtime without the need to implement complex components such as a GC and concurrency.
Language Interoperability Support
With Enso aiming to be the be-all and end-all for the data-science world, the ability to seamlessly interoperate with other programming languages is key. This means that a user should be able to paste in some Python or R code and have it work properly.
From a simple perspective, there were no other options in this category. While the JVM would allow for interoperability with other JVM languages such as Scala, Kotlin, and Java itself, the two 'most important' languages for interoperation had no support. LLVM's story is similar, allowing users to use LLVM IR as a common interoperation format, but this is far less practical than the JVM. With GHC, any interoperation would have to be developed from-scratch and by hand, essentially ruling it out in this category.
With GraalVM supporting not only our primary interoperability targets, but also the whole JVM ecosystem and any language that targets LLVM, it is an absolute dream for ensuring that Enso can seamlessly communicate with a whole host of other programming languages.
Implementation Performance
Data science often involves the manipulation of very large amounts of data, and ensuring that an interactive environment like Enso doesn't slow down as it does so requires a high level of performance.
GraalVM's partially-evaluated-interpreter based approach allows the developers to write a 'naive interpreter' and automatically have the platform provide better performance. This is a stark contrast to all of the other listed options, each of which would require significant complexity around generating the right intermediate representation structures, as well as significant work on front-end language optimisations.
In essence, GraalVM provides for the best performance with the smallest amount of effort, while still providing comprehensive facilities to improve performance further in the future.
Maintenance Burden
Just as important as getting a working runtime is the ability for the developers to improve and evolve it. This encompasses many factors, but Enso is primarily concerned with being able to evolve without having to account for undue changes to the runtime.
LLVM provides a relatively stable IR target, so the maintenance burden wouldn't have been too onerous. Similarly for the JVM, where the bytecode format has been stable for many years. Though both projects add new instructions, they very rarely remove them, meaning that Enso's potential code generator would be able to work as the underlying platform evolves.
As mentioned before, however, the intermediate representations in GHC that Enso would have used as a target are very much changeable. This is due to their primary existence being to support GHC's version of Haskell, which means that they change often. Furthermore, their generation would require copying of many of the idiosyncrasies of GHC's lowering mechanisms, and in all likelihood place a significant burden on Enso's developers.
GraalVM, on the other hand, provides a stable interface to writing an interpreter that is far higher level than any of the other options. This API is very unlikely to change, but even if it does the high-level nature means that the maintenance burden of coping with those changes is significantly reduced. Furthermore, GraalVM comes with the truffle toolkit for building interpreters, and as a result provides many of the facilities required by Enso for free or at least for little effort.
The Runtime Components
Like any sensible large software project, Enso's runtime is modular and broken down into components. These are described in detail below.
Language Server
The language server component is responsible for controlling the runtime itself. It communicates with other portions of the ecosystem (such as the REPL and the Enso Studio backend) via a protocol. While this protocol is based on the Language Server Protocol, it has been extended significantly to better support Enso's use-cases.
Filesystem Driver
This component of the runtime deals with access from the runtime to external devices. This includes the Enso code files on disk, but is also responsible for watching filesystem resources (such as databases, files, and sockets) that are used by Enso programs.
Typechecker
The typechecker is the portion of the runtime that handles the type-inference and type-checking of Enso code. This is a sophisticated piece of machinery, with the primary theory under which it operates being described in the specification of the type system.
Optimiser
With much of Enso's performance relying on the JIT optimiser built into Graal, the native language optimiser instead relies on handling more front-end specific optimisations.
Interpreter and JIT
The interpreter component is responsible for the actual execution of Enso code. It is built on top of the Truffle framework provided by GraalVM, and is JIT compiled by GraalVM.
Cross-Cutting Concerns
The runtime also has to deal with a number of concerns that don't fit directly into the above components, but are nevertheless important parts of the design.
Caching
The runtime cache for Enso is a key part of how it delivers exceptional performance when working on big data sets. The key recognition, as seen in many data processing tools, is that changing code or data often doesn't require the interpreter to recompute the entire program. Instead, it can only recompute the portions that are required of it, while using cached results for the rest.
Profiling and Debugging
Similarly important to the Enso user experience is the ability to visually debug and profile programs. This component deals with the retrieval, storage, and manipulation of profiling data, as well as the ability to debug programs in Enso using standard and non-standard debugging paradigms.
Foreign Language Interoperability
This component deals with using the GraalVM language interoperability features to provide a seamless interface to foreign code from inside Enso.
Lightweight Concurrency
Though not strictly a component, this section deals with how Enso can provide its users with lightweight concurrency primitives in the form of green threads.
The Initial Version of the Runtime
In order to have a working version of the new runtime as quickly as possible, it was decided to design and build an initial, stripped-down version of the final design. This design focused on development of a minimal working subset of the runtime that would allow Enso to run.
Development Considerations
As part of developing the new Enso runtime, the following things need to be accounted for. This is to ensure that the eventual quality of the software is high, and that we also provide a product that is actually useful to our users.
- Benchmarking: A comprehensive micro and macro benchmark suite that tests all the components of the runtime. This should be accompanied by a regression suite to catch performance regressions.
- Execution Tests: A test suite that checks that executing Enso programs results in the correct outputs.
- Typechecker Tests: A test suite that ensures that changes made to the typechecker do not result in acceptance of ill-typed programs, or rejection of well-typed programs.
- Caching Tests: A test suite that ensures that data is evicted from the cache when it should be, and retained when it should be.