Nearly zero-overhead input/output streams for Nim
Go to file
Jacek Sieka cf8d4d2263
restore 1.6 compat (#61)
2024-10-29 17:20:30 +01:00
.github/workflows restore 1.6 compat (#61) 2024-10-29 17:20:30 +01:00
faststreams restore 1.6 compat (#61) 2024-10-29 17:20:30 +01:00
tests fix ensureRunway after delayFixedSizeWrite (fixes #28) (#51) 2024-03-09 08:28:32 +01:00
.gitignore ignore build 2022-11-23 16:29:42 +01:00
LICENSE-APACHEv2 Initial version (extracted from nim-serialization) 2018-12-17 23:49:42 +02:00
LICENSE-MIT Initial version (extracted from nim-serialization) 2018-12-17 23:49:42 +02:00
README.md Fix spelling & grammer errors in the README (#53) 2024-04-03 07:27:47 +02:00
config.nims Nimble lock file 2022-04-12 19:24:02 +03:00
faststreams.nim Implement the async pipelines and fsMultiSync 2020-05-05 02:49:39 +03:00
faststreams.nimble update ci.yml and be more explicit in .nimble (#55) 2024-06-26 02:12:29 +00:00
nim.cfg move `-d:nimRawSetjmp` to config.nims (#30) 2022-09-17 08:53:56 +03:00

README.md

nim-faststreams

License: Apache License: MIT Stability: experimental Github action

FastStreams is a highly efficient library for all your I/O needs.

It offers nearly zero-overhead synchronous and asynchronous streams for handling inputs and outputs of various types:

  • Memory inputs and outputs for serialization frameworks and parsers
  • File inputs and outputs
  • Pipes and Process I/O
  • Networking

The library aims to provide a common interface between all stream types that allows the application code to be easily portable to different back-end event loops. In particular, Chronos and AsyncDispatch are already supported. It's envisioned that the library will also gain support for the Nginx event loop to allow the creation of web applications running as Nginx run-time modules and the SeaStar event loop for the development of extremely low-latency services taking advantage of kernel-bypass networking.

What does zero-overhead mean?

Even though FastStreams support multiple stream types, the API is designed in a way that allows the read and write operations to be handled without any dynamic dispatch in the majority of cases.

In particular, reading from a memoryInput or writing to a memoryOutput will have similar performance to a loop iterating over an openArray or another loop populating a pre-allocated string. memFileInput offers the same performance characteristics when working with files. The idiomatic use of the APIs with the rest of the stream types will result in a highly efficient memory allocation patterns and zero-copy performance in a great variety of real-world use cases such as:

  • Parsers for data formats and protocols employing formal grammars
  • Block ciphers
  • Compressors and decompressors
  • Stream multiplexers

The zero-copy behavior and low-memory usage is maintained even when multiple streams are layered on top of each other while back-pressure is properly accounted for. This makes FastStreams ideal for implementing highly-flexible networking stacks such as LibP2P.

The key ideas in the FastStreams design

FastStreams is heavily inspired by the System.IO.Pipelines API which was developed and released by Microsoft in 2018 and is considered the result of multiple years of evolution over similar APIs shipped in previous SDKs.

We highly recommend reading the following two articles which provide an in-depth explanation for the benefits of the design:

Here, we'll only summarize the main insights:

Obtaining data from the input device is not the same as consuming it.

When protocols and formats are layered on top of each other, it's highly inconvenient to handle a read operation that can return an arbitrary amount of data. If not enough data was returned, you may need to copy the available bytes into a local buffer and then repeat the reading operation until enough data is gathered and the local buffer can be processed. On the other hand, if more data was received, you need to complete the current stage of processing and then somehow feed the remaining bytes into the next stage of processing (e.g. this might be a nested format or a different parsing branch in the formal grammar of the protocol). Both of these scenarios require logic that is difficult to write correctly and results in unnecessary copying of the input bytes.

A major difference in the FastStreams design is that the arbitrary-length data obtained from the input device is managed by the stream itself while you are provided with an API allowing you to control precisely how much data is consumed from the stream. Consuming the buffered content does not invoke costly asynchronous calls and you are allowed to peek at the stream contents before deciding which step to take next (something crucial for handling formal grammars). Thus, using the FastStreams API results in code that is both highly efficient and easy to author.

Higher efficiency is possible if we say goodbye to the good old single buffer.

The buffering logic inside the stream divides the data into "pages" which are allocated with a known fast path in the Nim allocator and which can be efficiently transferred between streams and threads in the layered streams scenario or in IPC mechanisms such as AsyncChannel. The consuming code can be aware of this, but doesn't need to. The most idiomatic usage of the API handles the buffer switching logic automatically for the user.

Nevertheless, the buffering logic can be configured for unbuffered reads and writes and it supports efficiently various common real-world patterns such as:

  • Length prefixes

    To handle protocols with length prefixes without any memory overhead, the output streams support "delayed writes" where a portion of the stream content is specified only after the prefixed content is written to the stream.

  • Block compressors and Block ciphers

    These can benefit significantly from a more precise control over the stride of the buffered pages which can be configured to match the block size of the encoder.

  • Content with known length

    Some streams have a known length which allows us to accurately estimate the size of the transformed content. The len and ensureRunway APIs make sure such cases are handled as optimally as possible.

Basic API usage

The FastStreams API consists of ony few major object types:

InputStream

An InputStream manages a particular input device. The library offers out of the box the following input stream types:

  • fileInput

    For reading files through the familiar fread API from the C run-time.

  • memFileInput

    For reading memory mapped files which provides the best performance.

  • unsafeMemoryInput

    For handling strings, sequences and openarrays as an input stream.
    You are responsible for ensuring that the backing buffer won't be invalidated while the stream is being used.

  • memoryInput

    Primarily used to consume the contents written to a previously populated output stream, but it can also be used to consume the contents of strings and sequences in a memory-safe way (by creating a copy).

  • pipeInput (async)

    For arbitrary communication between a producer and a consumer.

  • chronosInput (async)

    Enabled by importing faststreams/chronos_adapters.
    It can represent any Chronos Transport as an input stream.

  • asyncSocketInput (async)

    Enabled by importing faststreams/std_adapters.
    Allows using Nim's standard library AsyncSocket type as an input stream.

You can extend the library with new InputStream types without modifying it. Please see the inline code documentation of InputStreamVTable for more details.

All of the above APIs are possible constructors for creating an InputStream. The stream instances will manage their resources through destructors, but you might want to close them explicitly in async context or when you need to handle the possible errors from the closing operation.

Here is an example usage:

import
  faststreams/inputs

var
  jsonString = "[1, 2, 3]"
  jsonNodes = parseJson(unsafeMemoryInput(jsonString))
  moreNodes = parseJson(fileInput("data.json"))

The example above assumes we might have a parseJson function accepting an InputStream. Here how this function could be defined:

import
  faststreams/inputs

proc scanString(stream: InputStream): JsonToken {.fsMultiSync.} =
  result = newStringToken()

  advance stream # skip the opening quote

  while stream.readable:
    let nextChar = stream.read.char
    case nextChar
    of '\'':
      if stream.readable:
        let escaped = stream.read.char
        case escaped
        of 'n': result.add '\n'
        of 't': result.add '\t'
        else: result.add escaped
      else:
        error(UnexpectedEndOfFile)
    of '"'
      return
    else:
      result.add nextChar

  error(UnexpectedEndOfFile)

proc nextToken(stream: InputStream): JsonToken {.fsMultiSync.} =
  while stream.readable:
    case stream.peek.char
    of '"':
      result = scanString(stream)
    of '0'..'9':
      result = scanNumber(stream)
    of 'a'..'z', 'A'..'Z', '_':
      result = scanIdentifier(stream)
    of '{':
      advance stream # skip the character
      result = objectStartToken
    ...

  return eofToken

proc parseJson(stream: InputStream): JsonNode {.fsMultiSync.} =
  while (let token = nextToken(stream); token != eofToken):
    case token
    of numberToken:
      result = newJsonNumber(token.num)
    of stringToken:
      result = newJsonString(token.str)
    of objectStartToken:
      result = parseObject(stream)
    ...

The above example is nothing but a toy program, but we can already see many usage patterns of the InputStream type. For a more sophisticated and complete implementation of a JSON parser, please see the nim-json-serialization package.

As we can see from the example above, calling stream.read should always be preceded by a call to stream.readable. When the stream is in the readable state, we can also peek at the next character before we decide how to proceed. Besides calling read, we can also mark the data as consumed by calling stream.advance.

The above APIs demonstrate how you can consume the data one byte at the time. Common wisdom might tell you that this should be inefficient, but that's not the case with FastStreams. The loop while stream.readable: stream.read will compile to very efficient inlined code that performs nothing more than pointer increments and comparisons. This will be true even when working with async streams.

The readable check is the only place where our code may block (or await). Only when all the data in the stream buffers have been consumed, the stream will invoke a new read operation on the backing input device and this may repopulate the buffers with an arbitrary number of new bytes.

Sometimes, you need to check whether the stream contains at least a specific number of bytes. You can use the stream.readable(N) API to achieve this.

Reading multiple bytes at once is then possible with stream.read(N), but if you need to store the bytes in an object field or another long-term storage location, consider using stream.readInto(destination) which may result in zero-copy operation. It can also be used to implement unbuffered reading.

AsyncInputStream and fsMultiSync

An astute reader might have wondered what is the purpose of the custom pragma fsMultiSync used in the examples above? It is a simple macro generating an additional async copy of our stream processing functions where all the input types are replaced by their async counterparts (e.g. AsyncInputStream) and the return type is wrapped in a Future as usual.

The standard API of InputStream and AsyncInputStream is exactly the same. Operations such as readable will just invoke await behind the scenes, but there is one key difference - the await will be triggered only when there is not enough data already stored in the stream buffers. Thus, in the great majority of cases, we avoid the high cost of instantiating a Future and yielding control to the event loop.

We highly recommend implementing most of your stream processing code through the fsMultiSync pragma. This ensures the best possible performance and makes the code more easily testable (e.g. with inputs stored on disk). FastStreams ships with a set of fuzzing tools that will help you ensure that your code behaves correctly with arbitrary data and/or arbitrary interruption points.

Nevertheless, if you need a more traditional async API, please be aware that all of the functions discussed in this README also have an *Async suffix form that returns a Future (e.g. readableAsync, readAsync, etc).

One exception to the above rule is the helper stream.timeoutToNextByte(t) which can be used to detect situations where your communicating party is failing to send data in time. It accepts a Duration or an existing deadline Future and it's usually used like this:

proc performHandshake(c: Connection): bool {.async.} =
  if c.inputStream.timeoutToNextByte(HANDSHAKE_TIMEOUT):
    # The other party didn't send us anything in time,
    # We close the connection:
    close c
    return false

  while c.inputStream.readable:
    ...

It is assumed that in traditional async code, timeouts will be managed more explicitly with sleepAsync and the or operator defined over futures.

Range-restricted reads

Protocols transmitting serialized payloads often provide information regarding the size of the payload. When you invoke the deserialization routine, it's preferable if the provided boundaries are treated like an "end of file" marker for the deserializer. FastStreams provides an easy way to achieve this without extra copies and memory allocations through the withReadableRange facility. Here is a typical usage:

proc decodeFrame(s: AsyncInputStream, DecodedType: type): Option[DecodedType] =
  if not s.readable(4):
    return

  let lengthPrefix = toInt32 s.read(4)
  if s.readable(lengthPrefix):
    s.withReadableRange(lengthPrefix, range):
      range.readValue(Json, DecodedType)

Please note that the above example uses the nim-serialization library

Simply, inside the withReadableRange block, range becomes a stream for which s.readable will return false as soon as the Json parser has consumed the specified number of bytes.

Furthermore, withReadableRange guarantees that all stream operations within the block will be non-blocking, so it will transform the AsyncInputStream into a regular InputStream. Depending on the complexity of the stream processing functions, this will often lead to significant performance gains.

OutputStream and AsyncOutputStream

An OutputStream manages a particular output device. The library offers out of the box the following output stream types:

  • writeFileOutput

    For writing files through the familiar fwrite API from the C run-time.

  • memoryOutput

    For building a string or a seq[byte] result.

  • unsafeMemoryOutput

    For writing to an arbitrary existing buffer.
    You are responsible for ensuring that the backing buffer won't be invalidated while the stream is being used.

  • pipeOutput (async)

    For arbitrary communication between a produced and a consumer.

  • chronosOutput (async)

    Enabled by importing faststreams/chronos_adapters.
    It can represent any Chronos Transport as an input stream.

  • asyncSocketOutput (async)

    Enabled by importing faststreams/std_adapters.
    Allows using Nim's standard library AsyncSocket type as an output stream.

You can extend the library with new OutputStream types without modifying it. Please see the inline code documentation of OutputStreamVTable for more details.

All of the above APIs are possible constructors for creating an OutputStream. The stream instances will manage their resources through destructors, but you might want to close them explicitly in async context or when you need to handle the possible errors from the closing operation.

Here is an example usage:

import
  faststreams/outputs

type
  ABC = object
    a: int
    b: char
    c: string

var stream = memoryOutput()
stream.writeNimRepr(ABC(a: 1, b: 'b', c: "str"))
var repr = stream.getOutput(string)

The writeNimRepr in the above example is not part of the library, but let's see how it can be implemented:

import
  typetraits, faststreams/outputs

proc writeNimRepr*(stream: OutputStream, str: string) =
  stream.write '"'

  for c in str:
    if c == '"':
      stream.write ['\'', '"']
    else:
      stream.write c

  stream.write '"'

proc writeNimRepr*(stream: OutputStream, x: char) =
  stream.write ['\'', x, '\'']

proc writeNimRepr*(stream: OutputStream, x: int) =
  stream.write $x # Making this more optimal has been left
                  # as an exercise for the reader

proc writeNimRepr*[T](stream: OutputStream, obj: T) =
  stream.write typetraits.name(T)
  stream.write '('

  var firstField = true
  for name, val in fieldPairs(obj):
    if not firstField:
      stream.write ", "

    stream.write name
    stream.write ": "
    stream.writeNimRepr val

    firstField = false

  stream.write ')'

When the stream is created, its output buffers will be initialized with a single page of pageSize bytes (specified at stream creation). Calls to write will just populate this page until it becomes full and only then it would be sent to the output device.

As the example demonstrates, a memoryOutput will continue buffering pages until they can be finally concatenated and returned in stream.getOutput. If the output fits within a single page, it will be efficiently moved to the getOutput result. When the output size is known upfront you can ensure that this optimization is used by calling stream.ensureRunway before any writes, but please note that the library is free to ignore this hint in async context or if a maximum memory usage policy is specified.

In a non-memory stream, any writes larger than a page or issued through the writeNow API will be sent to the output device immediately.

Please note that even in async context, write will complete immediately. To handle back-pressure properly, use stream.flush or stream.waitForConsumer which will ensure that the buffered data is drained to a specified number of bytes before continuing. The rationale here is that introducing an interruption point at every write produces less optimal code, but if this is desired you can use the stream.writeAndWait API.

If you have existing algorithms that output data to an openArray, you can use the stream.getWritableBytes API to continue using them without introducing any intermediate buffers.

Delayed Writes

Many protocols and formats employ fixed-size and variable-size length prefixes that have been traditionally difficult to handle because they require you to either measure the size of the content before writing it to the stream, or even worse, serialize it to a memory buffer in order to determine its size.

FastStreams supports handling such length prefixes with a zero-copy mechanism that doesn't require additional memory allocations. stream.delayFixedSizeWrite and stream.delayVarSizeWrite are APIs that return a WriteCursor object that can be used to implement a delayed write to the stream. After obtaining the write cursor you can take a note of the current pos in the stream and then continue issuing stream.write operations normally. After all of the content is written, you obtain pos again to determine the final value of the length prefix. Throughout the whole time, you are free to call write on the cursor to populate the "hole" left in the stream with bytes, but at the end you must call finalize to unlock the stream for flushing. You can also perform the finalization in one step with finalWrite (the one-step approach is mandatory for variable-size prefixes).

Pipeline

(This section is a stub and it will be expanded with more details in the future)

A Pipeline represents a chain of transformations that should be applied to a stream. It starts with an InputStream followed by one or more transformation steps and ending with a result.

Each transformation step is a function of the kind:

type PipelineStep* = proc (i: InputStream, o: OutputStream)
                          {.gcsafe, raises: [Defect, CatchableError].}

A result obtaining operation is a function of the kind:

type PipelineResultProc*[T] = proc (i: InputStream): T
                                   {.gcsafe, raises: [Defect, CatchableError].}

Please note that stream.getOutput is an example of such a function.

Pipelines executed in place with executePipeline API. If the first input source is async, then the whole pipeline with be executing asynchronously which can result in a much lower memory usage.

The pipeline transformation steps are usually employing the fsMultiSync pragma to make them usable in both synchronous and asynchronous scenarios.

Please note that the above higher-level APIs are just about simplifying the instantiation of multiple Pipe objects that can be used to hook input and output streams in arbitrary ways.

A stream multiplexer for example is likely to rely on the lower-level Pipe objects and the underlying PageBuffers directly.

License

Licensed and distributed under either of

or

at your option. This file may not be copied, modified, or distributed except according to those terms.