Tether: remote messaging between Smalltalks with WebSockets

In my previous post, I introduced a new topology for distributed computation with Smalltalk: an object memory in SqueakJS in a web browser, paired by remote-messaging connection with another object memory in Cog in a native app, and connected with other SqueakJS/Cog pairs on other physical machines. The remote-messaging protocol that the memories speak is called Tether. I’ll go into a few details here.

passive frame-based messaging with WebSockets

We begin with a major constraint imposed by running in a web browser: we’re sandboxed, unable to listen for network connections. We must initiate a remote-messaging conversation by connecting to a listening server on a traditional operating system. Over the last few years, the W3C WebSockets standard has received widespread support in every major web browser. We can rely on the ability to create JavaScript WebSocket objects with the SqueakJS JavaScript bridge, and we can easily implement the WebSocket API in Smalltalk on non-web platforms using normal TCP sockets.

WebSockets use callback functions to deliver messages, or frames of bytes, between conversants. The Tether protocol imposes a structure on those bytes, which are processed on each side of the connection by instances of class Tether. The first four bytes are a 32-bit tag which indicates the Smalltalk class which should interpret the rest of the frame. In the case of a remote message, this is class Tether. Successive bytes indicate the message selector to perform, the receiver of the message, and the message parameters.

The message receiver is expressed as a 32-bit key into a table of exposed objects, maintained by the Tether instance handling the connection. The Tether instances themselves expose their identities to each other at the beginning of the conversation. Objects that aren’t specified by reference to an exposed-object table (such as message selectors) are expressed through serialization.

live serialization

The fact that both sides of the conversation are objects in live Smalltalk systems affords many optimizations that aren’t possible when serializing objects to a static file. For this reason, I call this live serialization. For example, when transferring a compiled method between systems, if the method’s literals are objects which already exist in the receiving system, we may write references to them rather than serializing them.

We can also take special measures when the receiving system is missing the classes whose instances we want to transfer. Instead of assuming in advance that the receiving system lacks the classes whose instances we’re transferring, and including them in our payload, as a static serialization file would, we can transfer such classes only on demand. This yields much higher accuracy in object transfer, and far fewer bytes sent over the wire. Since live serialization is part of a complete remote messaging protocol, any messages at all can be sent from either side to complete an object transfer.

With a receiver, selector, and parameters specified, the receiving system can perform the message sent from the sending system. Each object in the system is responsible for serializing itself over a Tether. If the answer to the remote message is a literal object, like a symbol or integer, it will write the bytes of its value on the Tether instance handling the message. If the answer isn’t a literal object, by default it will write a reference to itself. Developers can choose to pass objects by value or by reference as they see fit.

scheduling

Tether performs every remote message in a distinct process, so that no system blocks waiting for an answer to be sent back over the network. Each remote message-send is assigned a unique identifier, and each answer is sent with the ID of its message-send as metadata, so that it can be delivered to the correct waiting process.

This scheme extends the traditional imperative messaging semantics of Smalltalk to any number of machines, and each message-send may involve receiver and parameter objects which are all on different machines. Every message-send may invoke any number of further remote messages before answering.

transparent proxies

An object which represents an object on a remote system is called a proxy. Ideally, it forwards every message sent to it to the remote object, and so provides the illusion of transparent remote messaging. Remote messaging in Smalltalk is often done by using a proxy class which inherits and implements as few messages as possible, and overriding the handler message sent by the virtual machine when a message is not understood. This provides enough coverage to do many useful things, but some messages handled specially by the virtual machine are not forwarded. Some use cases, like remote debugging, require forwarding even those special messages.

To achieve total forwarding coverage, we must modify the virtual machine. There are some situations where this is undesirable (e.g., a lack of tools or expertise, or a requirement to use a past virtual machine unmodified). Tether uses the “does not understand” tactic above in these situations, but provides a modified virtual machine for the rest. In this virtual machine, message forwarding is triggered during method lookup for instances of a specific proxy class (which can be located anywhere in the class hierarchy). Method caching and methods implemented directly as virtual machine instructions are also appropriately adapted. There are a few messages which proxies must understand locally, to participate in live serialization. These messages are also handled specially by the virtual machine.

see for yourself

Tether is an integral part of the Context project from Black Page Digital. A demo of remote messaging between SqueakJS and Cog is available. In tomorrow’s post, I’ll discuss an everyday application of remote messaging.

 

Advertisements

3 Responses to “Tether: remote messaging between Smalltalks with WebSockets”

  1. Hari Balaraman Says:

    Brilliant!

  2. […] « Tether: remote messaging between Smalltalks with WebSockets […]

  3. […] steps include getting the Tether remote messaging protocol and Snowglobe app streaming working between Pharo and Squeak, all running […]

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: