[jboss-jira] [JBoss JIRA] (JGRP-2218) New payload interface

[Sanne Grinovero (JIRA)]

    [ https://issues.jboss.org/browse/JGRP-2218?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13486154#comment-13486154 ] 

Sanne Grinovero commented on JGRP-2218:
---------------------------------------

Hi Bela,

{quote}but using an interface makes no difference compared to an abstract base class wrt memory footprint.{quote}

Remember our experiments with Java Object Layour in 2013 ?
 - http://lists.jboss.org/pipermail/infinispan-dev/2013-December/014311.html

You migtht be right as it might not matter, but it depends if you're able to keep the fields in the parent "packed". So I'd also question this choice, as even if you're able to do that today, it doesn't give you much flexibility into future improvements, e.g. being able to remove an unnecessary field from the parent would not help.


{quote}The UPerf test with 4 nodes shows slightly higher memory allocation for the payload use case ..
[...]
I assume that the higher allocation is caused by the messages needing more memory{quote}

Just checking, are you normalizing the memory allocation costs for the amount of messages being transferred? Just remember that if the new solutions sends more, it might not be surprising it consumes more.

BTW in JDK9 the flight recorder has been open sourced, massively improved and has an API: you might be able to write ANT tasks driving it in automation to produce verification reports, quite neat to use during such development but also for future pre-release verifications.

I'm unable to open the JFRs right now but I'll try next week. I'm interested to see if - even if it's possibly using more memory - if this is being consumed from cheaper regions, that would be an improvement!



> New payload interface
> ---------------------
>
>                 Key: JGRP-2218
>                 URL: https://issues.jboss.org/browse/JGRP-2218
>             Project: JGroups
>          Issue Type: Feature Request
>            Reporter: Bela Ban
>            Assignee: Bela Ban
>             Fix For: 5.0
>
>         Attachments: jgrp-2218.jfr, master.jfr
>
>
> h3. Goal
> Change payload in {{Message}} from byte[] arrays to a {{Payload}} interface which can have multiple implementations.
> h3. Reason
> Currently, having to pass a byte[] array to a message leads to unnecessary copying:
> * When an application has a ref to an NIO (direct) {{ByteBuffer}}, the bytes in the byte buffer have to be copied into a byte[] array and then set in the message
> * When the application sends around byte[] arrays, but also wants to add some additional metadata, e.g. type (1000-byte requests/responses), it needs to create a new byte[] array of (say) 1001 bytes and copy the data (1000 bytes) plus the request type (1 byte) into the new copy. Example: {{MPerf}} and {{UPerf}}
> * When an object has to be sent (e.g. in Infinispan), the object has to be marshalled into a byte[] array (first allocation) and then added to the message. With the suggested {{ObjectPayload}} (below), marshalling of the object would occur late, and it would be marshalled directly into the output stream of the bundler, eliminating the byte[] array allocation made by the application.
> h3. Design
> Instead of copying, the application creates an instance of {{Payload}} and sets the payload in {{Message}}. The {{Payload}} is then passed all the way down into the transport where it is marshalled and sent. There can be a number of payload implementations, e.g.
> * {{ByteArrayPayload}}: wraps a byte[] array with an offset and length
> * {{NioDirectPayload}}: wraps an NIO direct {{ByteBuffer}}
> * {{NioHeapPayload}}: wraps an NIO heap-based {{ByteBuffer}}
> * {{CompositePayload}}: wraps multiple Buffers. E.g. type (1 byte) and data (1000 bytes) as described above
> * {{IntPayload}}: a single integer
> * {{ObjectPayload}}: has an Object and a ClassLoader (for reading), plus a Marshaller which know how to marshal the object, this allows for objects to be passed in payloads and they're only marshalled at the end (transport).
> * {{PartialPayload}}: a ref to a {{Payload}}, with an offset and length
> * {{InputStreamPayload}}: has a ref to an input stream and copies data from input- to output stream when marshalling
> The {{Payload}} interface has methods:
> * {{size()}}
> * {{writeTo(DataOutput)}}
> * {{readFrom(DataInput)}}
> * {{getInput()}}: this provides a {{DataInput}} stream for reading from the underlying payload
> and  possibly also 
> * {{acquire()}} and 
> * {{release()}} (for ref-counting)
> * {{copy()}}
> Each payload impl has an ID and it should be possible to register new impls. A {{PayloadFactory}} maintains a mapping between IDs and impl classes.
> When marshalling a {{Payload}}, the ID is written first, followed by the payload's {{writeTo()}} method. When reading payloads, the {{PayloadFactory}} is used to create instances from IDs.
> h4. Fragmentation
> When fragmenting a buffer, the fragments are instances of {{PartialPayload}} which maintains an offset and length over an underlying payload. When marshalling a {{PartialPayload}}, only the part between offset and offset+length is written to the output stream.
> For fragmentation, method {{size()}} is crucial to determine whether a payload needs to be fragmented, or not. If, for example, a payload (e.g. an {{ObjectPayload}}) cannot determine the correct size, it may return {{-1}}. This leads to the {{ObjectPayload}} getting marshalled right away and getting wrapped into a {{ByteArrayPayload}}. So if {{size()}} cannot be determined, we have exactly the same behavior as what's currently done.
> h4. Reference counting
> If we implement ref-counting, then payloads can be reused as soon as the ref-count is 0. For example, when sending a message, the payload's ref-count could be incremented by the app calling {{acquire()}}. (Assuming the message is a unicast message), {{UNICAST3}} would increment the count to 2. This is needed because {{UNICAST3}} might have to retransmit the message if it was lost on the network, and meanwhile the payload cannot be reused (changed). The app calls {{release()}} when the {{JChannel.send()}} call returns, but the payload cannot be reused until {{UNICAST3}} calls {{release()}} as well. This will happen when an {{ACK}} for the given message has been received.
> h4. Payload factory
> When a request is received, the buffer is created from the bytes received on the network, based on the ID. This should be done by asking a {{PayloadFactory}} component for a new buffer. A naive implementation might create a new buffer every time, but a more sophisticated one might use a pool of payloads.
> The {{PayloadFactory}} instance could be replaced by one's own implementation; this allows for an application to control the lifecycle of payloads: thus the creation of buffers by the application and of payloads received over the network can be controlled by the same payload management impl.
> h4. Symmetry
> When sending a {{CompositePayload}} of a 500 byte {{ByteArrayPayload}} and a 1000 byte {{NioDirectPayload}}, would we want to also get the same {{CompositePayload}} consisting of 2 payloads on the receiver side, or would we want to combine the 2 payloads into one and make the 2 payloads refer to the same combined byte[] array (or NIO buffer)? Should this be made configurable?
> h4. ObjectPayload
> If ObjectPayload cannot determine the size of the serialized data, it should return {{-1}}. This means that {{Message.setPayload(ObjectPayload)}} would right away serialize {{ObjectPayload}} into {{ByteArrayPayload}}.
> This means we do have the {{byte[]}} array creation (same as now), but for object payloads which do implement {{size()}} correctly, we could still do late serialization.
> h5. ObjectPayload and fragmentation
> {{FRAG3}} could decorate {{ObjectPayload}} with a fragmentation payload, which generates fragments on serialization and sends them down the stack.
> h4. Misc
> * Since this issue includes API changes, the version will be 5.0



--
This message was sent by Atlassian JIRA
(v7.5.0#75005)