четверг, 21 июля 2016 г.

Major Seamless update

I glad to publish new version of Seamless (0.8.2) together with new doc.
It could be loaded by:

       Metacello new
           baseline: 'Seamless';
           repository: 'github://dionisiydk/Seamless';
           load

It works in Pharo 5 and 6. As usually feedback is welcome.

It is complete redesign of original version with the goal to make it more flexible, reliable and simple. (original version was created by Nikolaos Papoulias).

New version introduces new project Basys as foundation for new design.
Basys implements an abstract layer for networks which require bidirectional communication between clients and servers. (details here).
Seamless implements Basys network to organize transparent communication between distributed objects.

To use Seamless SeamlessNetwork should be created on client and server:

       network := SeamlessNetwork new.

To accept connections server should be started:

       network startServerOn: 40422.

Clients could connect to server and retrieve remote environment:

       remotePeer := network remotePeerAt: (TCPAddress ip: #[127 0 0 1] port: 40422).
       remoteSmalltalk := remotePeer remoteEnvironment.

       "or short version"
       remoteSmalltalk := network environmentAt: (TCPAddress localAt: 40422)

remoteSmalltalk here is proxy which delegates any received message to remote object. Remote messages are executed on server which return result back to client. Result can be returned as another proxy or as copy which contains another proxies.

In example result is reference to remote Smalltalk instance. It can access globals from remote environment:

       remoteTranscript := remoteSmalltalk at: #Transcript.
       remoteTranscript open; show: 'remote message'; cr

It will open transcript on server and print text on it.

Arguments of remote message are transferred to server with same logic as message result transferred to client. On server arguments can include proxies and server can send messages to them:

       remoteTranscript print: #(1 2 3 4 5)

Here array will be passed to server as reference. Then on server transcript will interact with it to print it. And as result client will receive messages from server.

Concrete transfer strategy is depends on transferred object. It is specified in method #seamlessDefaultTransferStrategy:

       Object>>seamlessDefaultTransferStrategy
            ^SeamlessTransferStrategy defaultByReference

       Number>>seamlessDefaultTransferStrategy
            ^SeamlessTransferStrategy defaultByValue

Default strategy could be overridden on network level:

       network transferByValue: (Instance of: Point).
       network transferByReference: (Identical to: #specificSymbol)

It allows to tune network for specific application to optimize communication between distributed objects. There are few other transfer strategies which allow minimize remote communication. Most interesting allows properties caching. It will transfer object reference together with specified properties. Following example will configure network to transfer class reference together with name:

       network transferByReference: (Kind to: Class) withCacheFor: #(name)

And then proxy for remote class will return #name from local cache instead of real remote call.

Previously announced project RemoteDebuggingTools uses all these feature to minimize communication between debugger and remote process. It allows to reduce number of remote messages for opening debugger from 800 to 13.

Now another important feature: block evaluation on remote peers:

       remotePeer evaluate: [1 + 2]. "==>3"

Given block is transferred to remote side and evaluated. Result is returned to client. As in other cases it could be proxy or normal object.

Block could use globals. On remote side they will be local globals of this remote environment. Following example will show notification on remote image:

       remotePeer evaluate: [Object inform: 'message from remote image'].

Temps and workspace variables can be used too. They will be transferred according to own strategies:

       | result |
       result := OrderedCollection new.
       remotePeer evaluate: [100 to: 500 do: [:i | result add: i factorial ]].

Non local return is also supported in regular Smalltalk semantics:

       remotePeer evaluate: [1 to: 10 do: [:i | i>5 ifTrue: [^i] ] ]. "==>6"

Also block could be evaluated asynchronously without waiting any result:

       | result |
       result := OrderedCollection new.
       remotePeer evaluateAsync: [result add: 1000 factorial]. "it will not wait result"

Seamless provides integration with GT tools. Remote proxies could be inspected to explore remote objects state with ability to execute remote scripts (doIt, printIt). It is shown on remote debugging demo.

To analyse remote communication Seamless implements special tool SeamlessLogger. It is explained in doc.  Here is little example for following code:

       remoteRect := remotePeer evaluate: [0@0 corner: 2@3].
       remoteRect area. "==>6".
       localRect := 0@0 corner: 5@2.
       remoteRect evaluate: [remoteRect area + localRect area] "==>16"

Statistics will show number of messages, receivers and bytes which was transferred over network in dimension of receiver class or message selector:


At the end I must say about two important issues which will be covered in next versions:

1) No garbage collection

SeamlessNetwork keeps all objects which was transferred by reference. They will never be cleaned while network is live. 
Now It could be done manually by evaluating "network destroy". It will clean all object caches and close all connections. It could be not safe because remote peers could use this objects. Seamless tries to handle it properly with clear errors in such cases. 
In future unused distributed objects will be cleaned automatically.

2) No security: no authorization and encryption. Now for security purpose you need external tools like VPN

Update: project was moved to github

Комментариев нет:

Отправить комментарий