Kresten
On Jul 7, 2005, at 6:00 PM, Hiram Chirino wrote:
Hi Kresten, On Jul 4, 2005, at 11:53 AM, Kresten Krab Thorup wrote:== first project ==Right now the Trifork ORB is using NIO for the server-side of IIOP, but "classic" IO for the client side. The NIO part is great because it lets us run all corba handling in a single selector thread backed by the appserver's thread pool. However, With the experience from working with this for the last 5 years, I would like to redo the core I/O subsystem, and so I have started to do the first steps towards this rework.The benefits of this, apart from cleaning up code that has grown over time, would be:- Reduce copying data through the stack. - Reduce thread usage further to support even more clients.- Off-load reading&writing - e.g. response writing to the framework so as to better handle slow clients.There are many reasons why I would like to do this, here is one: One optimization that we did at one point was to pool byte arrays, because the allocation of byte arrays (read: zeroing out memory) took way too much time. [I know - generally pooling in JVM's is a bad idea, but in this case it made a lot of sense] With this rewrite we would gain the same optimization one more time. CORBA Input/OutputStreams should be backed by java.nio.ByteBuffers directly, which will then be passed straight down to the NIO interface.The effort as I see it falls in two parts:- Asynchroneous I/O API (AIO). Based on the abstractions we have internally in the ORB, I'm doing a generalized version of something very similar to IBM's aio4j; future-based socket I/O. Unline aio4j however, the API runs straight on top of a Java SE 1.4 [no native code], and hooks into an external thread pool to provide the same API based on both NIO and "classic" IO technology.This area of work is absolutely fascinating for me. It seems that in ActiveMQ we have very similar requirements to your ORB. A standalone project was started to deal with exactly those issues at http://activeio.org. We built a flexible IO abstractions so that a client application can choose the API and implementation technology that best suites it's needs. Clients can choose to use a synchronous packet, asynchronous packet, or a synchronous stream interface. The And the client can choose what underlying IO technology will implement the chosen interface, it could choose to use standard blocking IO, or use NIO in blocking mode, or NIO in non-blocking mode, or IBM AIO, or JXTA sockets, or jvm pipes, etc. etc. All of this has already been implemented and is ASL 2.0 licensed and being used by Geronimo and ActiveMQ. Since you've been thinking this, it would be good to chat to see if it would suite your needs or if needs enhancing.Regards, Hiram- IIOP Streams on AIO. Based on the above, write InputStream/ OutputStream implementations, as well as connection management, backed by the AIO infrastructure and NIO direct buffers such that the underlying OS can stream data straight into the high-level structures.Kresten On Jul 4, 2005, at 2:37 PM, Geir Magnusson Jr. wrote:Joern, Thanks for the note. This is the right place to discuss. There are two separate threads of discussion that I can think of :1) Technical - if the donation fits technically into what we are doing (I'm sure it does...)2) Administrative - how, where and when As for #2 my questions to you are : a) What is the timing for this donation? How soon?b) Do you see this as coming straight to Geronimo to be part of Geronimo initially, or to the Apache Incubator where we could work on it with you and then make the decision of coming to the Geronimo project or being something else, like a stand-alone top- level project.c) I assume that you'd be offering committers to help us with the codebase and to continue working and expanding. Do you have an idea of how many?Thanks for doing this, and we look forward to discussion on both subjects above.geir -- Geir Magnusson Jr +1-203-665-6437 [EMAIL PROTECTED]
smime.p7s
Description: S/MIME cryptographic signature
