HTTP and Stream Processing

As we discussed in the previous post, we are thinking about a new design and implementation for the streams library in Open Dylan.

While the examples in this post are in Dylan and are using code from our HTTP server, these issues exist in HTTP frameworks in other languages. The code should be clear enough that little to no Dylan knowledge is required to understand the points being made here.

What does this have to do with HTTP? There are several pain points in our HTTP stack as it is currently written:

  • Requests are read in their entirety into memory, so a large request (such as a file upload) takes a significant amount of memory.
  • Responses often buffer their entire output in memory as well.
  • Because of the use of the existing streams library, we don't handle non-blocking sockets and require a thread per socket.
  • We don't have a good model for handling long-lasting connections such as might be used with Server Sent Events or WebSockets without tying up a thread for the duration of the socket being open.

We don't know yet what the new streams API will look like, but we can look at our ...

read more »

There are comments.

Beginning to Rethink Streams

Dylan's current streams library has served us moderately well over the years. However, it has some issues which can be addressed by a new design, expanding the range of problems for which it is suited.

How things are now

According to the current streams library's documentation, the design goals were:

  • A generic, easy-to-use interface for streaming over sequences and files. The same high-level interface for consuming or producing is available irrespective of the type of stream, or the types of the elements being streamed over.
  • Efficiency, especially for the common case of file I/O.
  • Access to an underlying buffer management protocol.

One of the things it was explicitly not designed to handle was, again, according to the design goals in the documentation:

  • A comprehensive range of I/O facilities for using memory-mapped files, network connections, and so on.

Unfortunately, the primary interface to our current network library is based on these very streams for which network connections were not a design goal. While this works in practice, it imposes some important limitations on our networking code. The biggest of these is that sockets can not be non-blocking as it is expected that reads and writes will complete ...

read more »

There are comments.

Integrating with LLDB

I spend a lot of time debugging Dylan code. Up until now, this has been a somewhat painful process when not using the IDE on Windows. (And I don't really use the IDE on Windows as it doesn't fit well into my workflow.) I finally reached the point where I decided that I wanted to improve our debugger integration.

Much of what is described below may be applicable to people working on debugging support for other languages.

Current State of Dylan Debugging

We have a debugger on Windows integrated with our IDE. This facility is not available on the other platforms that we support. There are many reasons for this:

  • The debug info that Open Dylan generates is only done on Windows.
  • The debugger code for interacting with the OS is only implemented for Windows.
  • The IDE itself is only available on Windows.

This means that debugging on Linux, FreeBSD and Mac OS X has traditionally been more challenging. We often resort to "printf debugging" and have some basic debug printing functions that can be invoked from the compiler so long as they don't crash. Debugging is really only effective with the C back-end as the HARP ...

read more »

There are comments.

Reducing Dispatch Overhead

Method dispatch in Dylan is the process by which the compiler and run-time collaborate to choose the right implementation of a function to call. This can get rather complex as a number of factors are involved in choosing the right method to call and whether that can be done at compile-time or deferred to run-time.

Unfortunately, full generic method dispatch at run-time in Open Dylan is currently not as fast as we would like. This means that when performance issues strike, they may well be due to the overhead of method dispatch.

Discussing and improving the overall performance of method dispatch isn't the subject of this post. That's going to require a fair bit of planning and work before that is resolved.

Instead, we're going to look at what to do when you're experiencing performance problems at particular call sites due to the overhead of method dispatch.

Sometimes, this is readily visible in the profiler:

/static/images/method_dispatch_in_profiler.png

Here, we can see that the percent-encode method in the uri library is going through dispatch to call member?. (I've translated from the names of the functions in C to their names in Dylan. The name mangling isn't that ...

read more »

There are comments.

Early Thoughts on Atom

I got a beta of the Atom editor soon after its initial release. The first thing that I decided to do was to add support for the Dylan language.

As we had a Textmate bundle available, this was an easy process:

apm init -p ~/.atom/packages/language-dylan \
  -c https://github.com/textmate/dylan.tmbundle

This was the start of a pleasurable experience with Atom. Creating and publishing a new package was easy and the command line tools were easy to use (and documented). After just a couple of minutes, I had a package that provided basic syntax highlighting and code folding for Dylan. It also worked on any Dylan file that I opened, unlike using our Textmate bundle in SublimeText which would occasionally hang on some files (while Textmate was fine). How's that for extensibility? :)

Next up, I decided to improve upon our new Dylan package. So I added support for auto-indenting and code snippets. This too was straight forward and easy.

As a comparison, I've also worked in the past with some other Dylan hackers on a plugin for IntelliJ to provide Dylan language support. While many things in IntelliJ are relatively easy to do, they all ...

read more »

There are comments.

Eliminating Copies in C-FFI

We've run into a situation fairly regularly when wrapping C libraries where we'd like to avoid making extra copies of data when passing it between Dylan and C code via the C-FFI. Related to this, we want to be able to use multiple types from the Dylan side of things such as <buffer>, <byte-vector> or <byte-string> without having to have separate code paths for each of them. Each of these classes can store raw byte data that we may want to share with other code.

An example that we'll work with in this discussion comes from bindings (in simplified and excerpted forms) for the LevelDB library:

extern void leveldb_put(
    leveldb_t* db,
    const leveldb_writeoptions_t* options,
    const char* key, size_t keylen,
    const char* val, size_t vallen,
    char** errptr);

<C-string>

The easiest way to bind leveldb_put would be to use <C-string>:

define C-function %leveldb-put
  input parameter db_ :: <leveldb-t*>;
  input parameter options_ :: <leveldb-writeoptions-t*>;
  input parameter key_ :: <C-string>;
  input parameter keylen_ :: <size-t>;
  input parameter val_ :: <C-string>;
  input parameter vallen_ :: <size-t>;
  output parameter errptr_ :: <char**>;
  c-name: "leveldb_put";
end;

However, this would mean that every key and value would have to be copied from an existing byte container (like <buffer>, <string> or <byte-vector>) to ...

read more »

There are comments.

Thoughts on a New Dylan Ecosystem

Dylan is in an unusual position for a programming language.

It has a pretty mature implementation, a specification, and a lot of high quality documentation. It has a long history and many people remember it from the early days of its history when it was being developed at Apple and other institutions.

On the other hand, the current community is fairly small as is the available set of libraries. This is typically a disadvantage, however it can be used to the community's advantage in a couple of ways.

Looking Forward

One way is that we can look forward and take advantage of what we know and want now without having to worry about backward compatibility.

A parallel might be seen with the early days of Node.js. Early on, Node.js was also working with a programming language with a specification and a mature implementation. However, Node.js was using JavaScript in a new way: to create server-side software employing asynchronous and non-blocking techniques. A big advantage that Node.js had was that there wasn't a large collection of libraries for doing server-side programming. The developers could (largely) start from a clean slate, producing libraries that fit in ...

read more »

There are comments.

A New IDE Approach

/static/images/DeftIDEA_smaller.png

We've previously written about the state of the Open Dylan IDE. In short, it is currently only available on Windows and looks like an application from the 1990s (because it is). Simply getting it running on Mac OS X or Linux would be a lot of work (an attempt was made to port the underlying GUI framework), not counting modernization and adding new features.

In late September, we had a seemingly crazy idea: What if we built a new IDE on an existing IDE framework. We've previously worked with Eclipse and didn't want to do that again. IntelliJ, however, officially supports IDEs for Java, Python, Ruby, PHP, and Objective C. Unofficially, it has support for a host of other languages including Erlang, Lua, Scala, HaXe and many others.

So, we wondered, what would it take to get some basics running for Dylan? In the end, the answer was that within 10 hours of first looking at IntelliJ plug-in authoring, we had it recognizing file types, doing some basic parsing for LID files and some other minor things.

This led to the creation of DeftIDEA, a new Dylan Environment For Tools running inside of IntelliJ IDEA.

Why did ...

read more »

There are comments.

Last Month's Hack-a-thon

Last month, we had our first hack-a-thon in many years. Attendance was spread out over the weekend and some great contributions got completed.

  • Carl Gay documented our regular expressions library.
  • Peter Housel made more solid progress on our LLVM compiler back-end.
  • Dustin Voss improved some aspects of our support for limited collections.
  • Ingo Promovicz improved support for locking I/O streams.
  • Bruce Mitchener worked on merging run-time libraries to help provide more consistent support for various features across the various compiler back-ends.
  • Francesco Ceccon worked on the gobject-introspection bindings and the generator for creating C-FFI bindings for the GTK+ family of libraries.
  • Hannes Mehnert addressed some issues in the networking library surrounding localhost connections.
  • Andreas Bogk returned from the dead and observed the goings on after a multi-year absence.
  • Tom Emerson began investigating providing support for Unicode in Dylan.
  • Robert Roland enhanced the sqlite bindings.

Thanks to everyone who participated and we're looking forward to our next hack-a-thon in a couple of months.

All of these changes and much much more will be present in our upcoming 2013.2 release!

read more »

There are comments.

Open Dylan's Compiler Back-ends

Open Dylan has 3 compiler back-ends currently:

  • HARP
  • C
  • LLVM

Harp

The HARP back-end generates native code for x86 processors and is used on Windows, FreeBSD and Linux. While for many years, this was our best and most mature compiler back-end, it is falling by the wayside and will be replaced by the LLVM back-end.

The issues with the HARP back-end are that it is 32-bit only and is more difficult to add support for things like new instructions, better scheduling, atomic operations and SSE support.

C

The C back-end generates C which is then compiled by either clang or gcc. We use the C back-end on Mac OS X, as well as Linux and FreeBSD on x86_64 processors.

The C back-end is relatively easy to work with and extend and is the easiest option for bringing up Open Dylan on a new platform. We are currently doing this with Linux on the ARM processor.

The C back-end used to generate code that was slower than HARP although in recent months, the C back-end has been optimized and many parts of it produce faster results than HARP.

The C back-end allows debugging the generated C (rather than at the Dylan ...

read more »

There are comments.