secc

Crates.iosecc
lib.rssecc
version0.0.10
sourcesrc
created_at2019-08-11 14:26:37.262029
updated_at2019-09-14 21:04:37.398209
descriptionImplements a Skip-Enabled Concurrent Channel (SECC) for Rust
homepagehttps://github.com/rsimmonsjr/secc
repositoryhttps://github.com/rsimmonsjr/secc
max_upload_size
id155830
size84,812
Robert Simmons Jr. (rsimmonsjr)

documentation

https://docs.rs/secc

README

Skip Enabled Concurrent Channel for Rust (SECC)

Latest version Build Status Average time to resolve an issue License

Description

An Skip Enabled Concurrent Channel (SECC) is a bounded capacity channel that supports multiple senders and multiple recievers and allows the receiver to temporarily skip receiving messages if they desire.

Messages in the channel need to be clonable to implement the [peek] functionality (which returns a clone of the message). For this reason it is advisable that the user chose a type that is efficiently clonable, such as an [Arc] to enclose a message that cannot be efficiently cloned.

The channel is a FIFO structure unless the user intends to skip one or more messages in which case a message could be read in a different order. The channel does, however, guarantee that the messages will remain in the same order as sent and, unless skipped, will be received in order.

SECC is implemented using two linked lists where one list acts as a pool of nodes and the other list acts as the queue holding the messages. This allows us to move nodes in and out of the list and even skip a message with O(1) efficiency. If there are 1000 messages and the user desires to skip one in the middle they will incur virtually the exact same performance cost as a normal read operation. There are only a couple of additional pointer operations necessary to remove a node out of the middle of the linked list that implements the queue. When a message is received from the channel the node holding the message is removed from the queue and appended to the tail of the pool. Conversely, when a message is sent to the channel the node moves from the head of the pool to the tail of the queue. In this manner nodes are constantly cycled in and out of the queue so we only need to allocate them once when the channel is created.

Examples

use secc::*;
use std::time::Duration;

let channel = create::<u8>(5, Duration::from_millis(10));
let (sender, receiver) = channel;
assert_eq!(Ok(()), sender.send(17));
assert_eq!(Ok(()), sender.send(19));
assert_eq!(Ok(()), sender.send(23));
assert_eq!(Ok(()), sender.send(29));
assert_eq!(Ok(17), receiver.receive());
assert_eq!(Ok(()), receiver.skip());
assert_eq!(Ok(23), receiver.receive());
assert_eq!(Ok(()), receiver.reset_skip());
assert_eq!(Ok(19), receiver.receive());

This code creates the channel and then sends it a series of messages. The first is received normally but then the user wants to skip the next message. The user can then receive in the middle of the channel, reset the skip and resume receiving normally.

What's New

  • 2019-09-13: 0.0.10
    • Issue #13: A Deadlock would occur if the timeout occurred while waiting for space or data.
    • BREAKING CHANGE Timeouts are in Duration objects now rather than milliseconds.
  • 2019-08-18: 0.0.9
    • Most unsafe code has been eliminated, enhancing stability.

Release Notes for All Versions

Design Principals

SECC was driven by the need for a multi-sender, multi-consumer channel that would have the ability to skip processing messages. There are many situation in which this is needed by a consumer such as the use case with Axiom where actors implement a finite state machine. That led me to go through many iterations of different designs until it became clear that a linked list was the only legitimate approach. The problem with a linked lists is that they typically burn a lot of CPU time in allocating new nodes on each enqueue. The solution was to use two linked lists, allocate all nodes up front and just logically move nodes around. The actual pointers to the next node or the various heads and tails are the indexes in the statically allocated slice of nodes. When send and receive operations happen, nodes are merely moved around logically but not physically.

Commit count: 27

cargo fmt