Notes on the Rust runtime 2013-02-02

There have been a number of discussions lately about fixing various aspects of the rust runtime, and over the next year it looks like we're going to be making a lot of changes that should simplify the implementation and improve performance.

'The Rust runtime', which I'll try to define later, provides a number of services with sometimes complex interactions, so there's a lot to think about when making big changes. Much of the code involved was written long before Rust was even self-hosting, in C++.

My personal long term goal is to have a well factored Rust runtime, with a simple and fast implementation written in Rust (typical Rust, not some massively unsafe dialect). I think we have the tools in the language to do this now. In the short term I'm going to be focusing on writing a task scheduler in Rust that is driven by a generic event loop, then make our non-blocking I/O as fast as possible by integrating it into the scheduler.

This post mostly for my own sake, organizing some thoughts I have on the subject, and collecting links to related issues and discussions.

Defining the current runtime

First I want to capture the broad architecture of the current runtime, then describe some of the problems it has.

Today, there is a library written in C++ called librustrt, and all crates depend on it. It contains a function called rust_start that begins execution of a Rust program by passing a Rust function pointer to a new instance of rust_kernel. rust_kernel is the runtime, and it manages a dynamic number of rust_schedulers, each of which in turn manages a fixed number of rust_sched_loops that schedule rust_tasks. No Rust code can run outside of a task.

When Rust code needs to access a runtime service it first aquires a pointer to its rust_task instance from thread-local storage. If it needs to it can get a rust_kernel pointer from the rust_task. The runtime exports an interface to Rust through C functions, mostly defined in rust_builtin.cpp and rust_upcall.cpp, most of which encapsulate the lookup of the task pointer.

Services provided by the runtime can be broadly categorized as 'task' services, and 'kernel' services. Task services have task-local effects and kernel services have kernel-global effects.

Kernel services:

Kernel services necessarily involve global thread synchronization, so are potential bottlenecks, and good candidates for removal and simplification.

Task services:

I also consider parts of core::pipes and core::task to be part of the runtime, because they contain a lot of private implementation details built on the services exported from librustrt. The GC (such as it is) is also written in Rust, but considered part of the runtime and depends on implementation details of the local heap. librustrt additionally includes some third party code that is not tied to rust_kernel and might or might not be considered part of the runtime. linenoise for instance is a line reading library used by std, and only lives in librustrt as a convenience.

Problems

Too much foreign code

The runtime is the only part of Rust not written in Rust. The interface between Rust and the Rust runtime consists of a handful of C functions taking opaque pointers, so can't take advantage of any of the Rust's nice features. The runtime could evolve much faster if it was written in Rust.

Task context

There are a lot of use cases for running Rust code outside of a Rust task, one of which is for writing the scheduler itself, but it's impossible to do anything without a *rust_task. We need a finer seperation of responsibilities.

Off-thread I/O

Non-blocking I/O must be done in a different thread. Under the current implementation we dispatch I/O requests to a global iotask, but this is thought to have a lot of overhead.

Locking bottlenecks

The current runtime is arranged as a heirarchy of entities, with the kernel coordinating schedulers that in turn manage a number of threads. Each level of this heirarchy contains one or more locks or atomic operations, and simply spawning a task hits several of them.

Stack management complexity

Most of the rust_task code is dedicated to managing stacks, even though the stack has little to do with tasks and scheduling. There are two completely separate code paths for extending the Rust stack ('stack growth') and running foreign code ('stack switching'), but these two use cases have very much in common.

Lifecycle management and internal complexity

It's quite difficult to trace through the relationship between the kernel, schedulers, scheduler threads and tasks. The conditions for triggering runtime shutdown and cleaning up the schedulers have been particularly hard to reason about. Linked failure, e.g. is still done by propagating a flag around task objects and checking for it in the runtime when it could be expressed through higher-level policies involving pipes. A lot of the complexity is for historical reasons and exacerbated by the inexpressiveness of interfaces between foreign code and Rust.

Task migration

Tasks never migrate between threads.

Logging

Rust's logging implementation is very old and lives in the C++ runtime. It should be completely rewritten to operate at a much higher level, using tasks and pipes.

Redefining the Rust runtime

In order to write the scheduler in Rust we need to be able to execute Rust code outside of a Rust task. Doing this requires some reconsideration of what the runtime is, in particular we need access to some services globally. Here is how I am starting to think of the runtime and it's components, even if this doesn't reflect current reality.

In the new regime core is the Rust runtime and librustrt doesn't exist. Some runtime services are accessible globally, the exchange heap being the most important. Another might be a global fallback console logging service. Some features may need to detect whether they are running inside or outside task context and change their behavior to use either a task, kernel, or global service as appropriate. For example I have a change that makes the FFI work either with or without a task.

Some day I would like all these services to be organized as such (global, kernel, task, etc.), all documented and living in intuitive places instead of being scattered all over core.

Current work

I've recently deleted the old message passing system, oldcomm, from the tree, and that took with it a big chunk of runtime code. With that out of the way I feel prepared to start prototyping a new scheduler.

I intend to begin by creating a very simple single-threaded Scheduler that schedules Tasks and is driven by a generic EventLoop trait implemented with uv. Once the scheduler is properly creating and scheduling tasks then I will start experimenting with integrating I/O. My goal will be to make the single-threaded case very fast, capture that performance in benchmarks, then extend that work to groups of work-stealing schedulers. The new code will be scheduler-centric, with one scheduler per thread, instead of schedulers being a group of scheduler threads. And instead of having a heirarchy of entities (kernel -> scheduler -> scheduler_thread -> task) I will aim have just a federation of schedulers, some of which steal tasks from each other. In the process of rebuilding the scheduler from the ground up I hope to be able to discard some of the assumptions made by the previous, ending up with something simpler.

I intend to write the scheduler itself using mostly typical Rust abstractions, so pretty soon I will want to port pipes to pthreads. For the single threaded case they won't be needed though - only once schedulers need to talk to each other will it matter.

I'm still thinking about how I want these runtime services to be represented and organized, but as I work on the scheduler I will probably create some sort of RuntimeService type that makes it clear what runtime capabilities your Rust code has access to at any time.

Tasks on deck:

Links