C++ – How does libuv compare to Boost/ASIO

boostboost-asioc++libuv

I'd be interested in aspects like:

scope/features
performance
maturity

Best Solution

Scope

Boost.Asio is a C++ library that started with a focus on networking, but its asynchronous I/O capabilities have been extended to other resources. Additionally, with Boost.Asio being part of the Boost libraries, its scope is slightly narrowed to prevent duplication with other Boost libraries. For example, Boost.Asio will not provide a thread abstraction, as Boost.Thread already provides one.

On the other hand, libuv is a C library designed to be the platform layer for Node.js. It provides an abstraction for IOCP on Windows, kqueue on macOS, and epoll on Linux. Additionally, it looks as though its scope has increased slightly to include abstractions and functionality, such as threads, threadpools, and inter-thread communication.

At their core, each library provides an event loop and asynchronous I/O capabilities. They have overlap for some of the basic features, such as timers, sockets, and asynchronous operations. libuv has a broader scope, and provides additional functionality, such as thread and synchronization abstractions, synchronous and asynchronous file system operations, process management, etc. In contrast, Boost.Asio's original networking focus surfaces, as it provides a richer set of network related capabilities, such as ICMP, SSL, synchronous blocking and non-blocking operations, and higher-level operations for common tasks, including reading from a stream until a newline is received.

Feature List

Here is the brief side-by-side comparison on some of the major features. Since developers using Boost.Asio often have other Boost libraries available, I have opted to consider additional Boost libraries if they are either directly provided or trivial to implement.

                         libuv          Boost
Event Loop:              yes            Asio
Threadpool:              yes            Asio + Threads
Threading:              
  Threads:               yes            Threads
  Synchronization:       yes            Threads
File System Operations:
  Synchronous:           yes            FileSystem
  Asynchronous:          yes            Asio + Filesystem
Timers:                  yes            Asio
Scatter/Gather I/O^[1]:    no             Asio
Networking:
  ICMP:                  no             Asio
  DNS Resolution:        async-only     Asio
  SSL:                   no             Asio
  TCP:                   async-only     Asio
  UDP:                   async-only     Asio
Signal:
  Handling:              yes            Asio
  Sending:               yes            no
IPC:
  UNIX Domain Sockets:   yes            Asio
  Windows Named Pipe:    yes            Asio
Process Management:
  Detaching:             yes            Process
  I/O Pipe:              yes            Process
  Spawning:              yes            Process
System Queries:
  CPU:                   yes            no
  Network Interface:     yes            no
Serial Ports:            no             yes
TTY:                     yes            no
Shared Library Loading:  yes            Extension^[2]

_{1. Scatter/Gather I/O.}

_{2. Boost.Extension was never submitted for review to Boost. As noted here, the author considers it to be complete.}

Event Loop

While both libuv and Boost.Asio provide event loops, there are some subtle differences between the two:

While libuv supports multiple event loops, it does not support running the same loop from multiple threads. For this reason, care needs to be taken when using the default loop (uv_default_loop()), rather than creating a new loop (uv_loop_new()), as another component may be running the default loop.
Boost.Asio does not have the notion of a default loop; all io_service are their own loops that allow for multiple threads to run. To support this Boost.Asio performs internal locking at the cost of some performance. Boost.Asio's revision history indicates that there have been several performance improvements to minimize the locking.

Threadpool

libuv's provides a threadpool through uv_queue_work. The threadpool size is configurable via the environment variable UV_THREADPOOL_SIZE. The work will be executed outside of the event loop and within the threadpool. Once the work is completed, the completion handler will be queued to run within the event loop.
While Boost.Asio does not provide a threadpool, the io_service can easily function as one as a result of io_service allowing multiple threads to invoke run. This places the responsibility of thread management and behavior to the user, as can be seen in this example.

Threading and Synchronization

libuv provides an abstraction to threads and synchronization types.
Boost.Thread provides a thread and synchronization types. Many of these types follow closely to the C++11 standard, but also provide some extensions. As a result of Boost.Asio allowing multiple threads to run a single event loop, it provides strands as a means to create a sequential invocation of event handlers without using explicit locking mechanisms.

File System Operations

libuv provides an abstraction to many file system operations. There is one function per operation, and each operation can either be synchronous blocking or asynchronous. If a callback is provided, then the operation will be executed asynchronously within an internal threadpool. If a callback is not provided, then the call will be synchronous blocking.
Boost.Filesystem provides synchronous blocking calls for many file system operations. These can be combined with Boost.Asio and a threadpool to create asynchronous file system operations.

Networking

libuv supports asynchronous operations on UDP and TCP sockets, as well as DNS resolution. Application developers should be aware that the underlying file descriptors are set to non-blocking. Therefore, native synchronous operations should check return values and errno for EAGAIN or EWOULDBLOCK.
Boost.Asio is a bit more rich in its networking support. In addition many of the features libuv's networking provides, Boost.Asio supporting SSL and ICMP sockets. Furthermore, Boost.Asio provides synchronous blocking and synchronous non-blocking operations, into addition to its asynchronous operations. There are numerous free standing functions that provide common higher-level operations, such as reading a set amount of bytes, or until a specified delimiter character is read.

Signal

libuv provides an abstraction kill and signal handling with its uv_signal_t type and uv_signal_* operations.
Boost.Asio does not provde an abstraction to kill, but its signal_set provides signal handling.

IPC

libuv abstracts Unix domain sockets and Windows named pipes through a single uv_pipe_t type.
Boost.Asio separates the two into local::stream_protocol::socket or local::datagram_protocol::socket, and windows::stream_handle.

API Differences

While the APIs are different based on the language alone, here are a few key differences:

Operation and Handler Association

Within Boost.Asio, there is a one-to-one mapping between an operation and a handler. For instance, each async_write operation will invoke the WriteHandler once. This is true for many of libuv operations and handlers. However, libuv's uv_async_send supports a many-to-one mapping. Multiple uv_async_send calls may result in the uv_async_cb being called once.

Call Chains vs. Watcher Loops

When dealing with task, such as reading from a stream/UDP, handling signals, or waiting on timers, Boost.Asio's asynchronous call chains are a bit more explicit. With libuv, a watcher is created to designate interests in a particular event. A loop is then started for the watcher, where a callback is provided. Upon receiving the event of interests, the callback will be invoked. On the other hand, Boost.Asio requires an operation to be issued each time the application is interested in handling the event.

To help illustrate this difference, here is an asynchronous read loop with Boost.Asio, where the async_receive call will be issued multiple times:

void start()
{
  socket.async_receive( buffer, handle_read ); ----.
}                                                  |
    .----------------------------------------------'
    |      .---------------------------------------.
    V      V                                       |
void handle_read( ... )                            |
{                                                  |
  std::cout << "got data" << std::endl;            |
  socket.async_receive( buffer, handle_read );   --'
}

And here is the same example with libuv, where handle_read is invoked each time the watcher observes that the socket has data:

uv_read_start( socket, alloc_buffer, handle_read ); --.
                                                      |
    .-------------------------------------------------'
    |
    V
void handle_read( ... )
{
  fprintf( stdout, "got data\n" );
}

Memory Allocation

As a result of the asynchronous call chains in Boost.Asio and the watchers in libuv, memory allocation often occurs at different times. With watchers, libuv defers allocation until after it receives an event that requires memory to handle. The allocation is done through a user callback, invoked internal to libuv, and defers deallocation responsibility of the application. On the other hand, many of the Boost.Asio operations require that the memory be allocated before issuing the asynchronous operation, such as the case of the buffer for async_read. Boost.Asio does provide null_buffers, that can be used to listen for an event, allowing applications to defer memory allocation until memory is needed, although this is deprecated.

This memory allocation difference also presents itself within the bind->listen->accept loop. With libuv, uv_listen creates an event loop that will invoke the user callback when a connection is ready to be accepted. This allows the application to defer the allocation of the client until a connection is being attempted. On the other hand, Boost.Asio's listen only changes the state of the acceptor. The async_accept listens for the connection event, and requires the peer to be allocated before being invoked.

Performance

Unfortunately, I do not have any concrete benchmark numbers to compare libuv and Boost.Asio. However, I have observed similar performance using the libraries in real-time and near-real-time applications. If hard numbers are desired, libuv's benchmark test may serve as a starting point.

Additionally, while profiling should be done to identify actual bottlenecks, be aware of memory allocations. For libuv, the memory allocation strategy is primarily limited to the allocator callback. On the other hand, Boost.Asio's API does not allow for an allocator callback, and instead pushes the allocation strategy to the application. However, the handlers/callbacks in Boost.Asio may be copied, allocated, and deallocated. Boost.Asio allows for applications to provide custom memory allocation functions in order to implement a memory allocation strategy for handlers.

Maturity

Boost.Asio

Asio's development dates back to at least OCT-2004, and it was accepted into Boost 1.35 on 22-MAR-2006 after undergoing a 20-day peer review. It also served as the reference implementation and API for Networking Library Proposal for TR2. Boost.Asio has a fair amount of documentation, although its usefulness varies from user to user.

The API also have a fairly consistent feel. Additionally, the asynchronous operations are explicit in the operation's name. For example, accept is synchronous blocking and async_accept is asynchronous. The API provides free functions for common I/O task, for instance, reading from a stream until a \r\n is read. Attention has also been given to hide some network specific details, such as the ip::address_v4::any() representing the "all interfaces" address of 0.0.0.0.

Finally, Boost 1.47+ provides handler tracking, which can prove to be useful when debugging, as well as C++11 support.

libuv

Based on their github graphs, Node.js's development dates back to at least FEB-2009, and libuv's development dates to MAR-2011. The uvbook is a great place for a libuv introduction. The API documentation is here.

Overall, the API is fairly consistent and easy to use. One anomaly that may be a source of confusion is that uv_tcp_listen creates a watcher loop. This is different than other watchers that generally have a uv_*_start and uv_*_stop pair of functions to control the life of the watcher loop. Also, some of the uv_fs_* operations have a decent amount of arguments (up to 7). With the synchronous and asynchronous behavior being determined on the presence of a callback (the last argument), the visibility of the synchronous behavior can be diminished.

Finally, a quick glance at the libuv commit history shows that the developers are very active.

Setting a bit

Use the bitwise OR operator (|) to set a bit.

number |= 1UL << n;

That will set the nth bit of number. n should be zero, if you want to set the 1st bit and so on upto n-1, if you want to set the nth bit.

Use 1ULL if number is wider than unsigned long; promotion of 1UL << n doesn't happen until after evaluating 1UL << n where it's undefined behaviour to shift by more than the width of a long. The same applies to all the rest of the examples.

Clearing a bit

Use the bitwise AND operator (&) to clear a bit.

number &= ~(1UL << n);

That will clear the nth bit of number. You must invert the bit string with the bitwise NOT operator (~), then AND it.

Toggling a bit

The XOR operator (^) can be used to toggle a bit.

number ^= 1UL << n;

That will toggle the nth bit of number.

Checking a bit

You didn't ask for this, but I might as well add it.

To check a bit, shift the number n to the right, then bitwise AND it:

bit = (number >> n) & 1U;

That will put the value of the nth bit of number into the variable bit.

Changing the nth bit to x

Setting the nth bit to either 1 or 0 can be achieved with the following on a 2's complement C++ implementation:

number ^= (-x ^ number) & (1UL << n);

Bit n will be set if x is 1, and cleared if x is 0. If x has some other value, you get garbage. x = !!x will booleanize it to 0 or 1.

To make this independent of 2's complement negation behaviour (where -1 has all bits set, unlike on a 1's complement or sign/magnitude C++ implementation), use unsigned negation.

number ^= (-(unsigned long)x ^ number) & (1UL << n);

unsigned long newbit = !!x;    // Also booleanize to force 0 or 1
number ^= (-newbit ^ number) & (1UL << n);

It's generally a good idea to use unsigned types for portable bit manipulation.

number = (number & ~(1UL << n)) | (x << n);

(number & ~(1UL << n)) will clear the nth bit and (x << n) will set the nth bit to x.

It's also generally a good idea to not to copy/paste code in general and so many people use preprocessor macros (like the community wiki answer further down) or some sort of encapsulation.

C++ – What does the explicit keyword mean

The compiler is allowed to make one implicit conversion to resolve the parameters to a function. What this means is that the compiler can use constructors callable with a single parameter to convert from one type to another in order to get the right type for a parameter.

Here's an example class with a constructor that can be used for implicit conversions:

class Foo
{
public:
  // single parameter constructor, can be used as an implicit conversion
  Foo (int foo) : m_foo (foo) 
  {
  }

  int GetFoo () { return m_foo; }

private:
  int m_foo;
};

Here's a simple function that takes a Foo object:

void DoBar (Foo foo)
{
  int i = foo.GetFoo ();
}

and here's where the DoBar function is called:

int main ()
{
  DoBar (42);
}

The argument is not a Foo object, but an int. However, there exists a constructor for Foo that takes an int so this constructor can be used to convert the parameter to the correct type.

The compiler is allowed to do this once for each parameter.

Prefixing the explicit keyword to the constructor prevents the compiler from using that constructor for implicit conversions. Adding it to the above class will create a compiler error at the function call DoBar (42). It is now necessary to call for conversion explicitly with DoBar (Foo (42))

The reason you might want to do this is to avoid accidental construction that can hide bugs.
Contrived example:

You have a MyString class with a constructor that constructs a string of the given size. You have a function print(const MyString&) (as well as an overload print (char *string)), and you call print(3) (when you actually intended to call print("3")). You expect it to print "3", but it prints an empty string of length 3 instead.