I already tried smth like that when tried to make marshalling parallel. This is how I tried to reapproach it. 1. allocate set of direct buffers per each session (connection). that was 1 large buffer, but sliced with smaller ones over large offheap array. 2. each thread acquires chunk by chunk from the session it wants to write to. 3. each chunk has a header - chunk length (31 bit) and last flag (1 bit) 4. message can span several chunks 5. all message chunks are passed to nio thread 6. nio thread gets chunks from the queue until queue is empty or total chunks length is equal to socket buffer and calls - java.nio.channels.SocketChannel#write(java.nio.ByteBuffer[], int, int)
On reader side 1. Read into large byte buffer 2. Scan it for chunks and messages that are fully read 3. Unmarshall each message in system pool passing slices over larger buffer This approach worked a bit slower than current solution, but brought a lot of complex code including more complex synchronizations. So, I decided not to continue with it. --Yakov 2016-10-14 9:09 GMT+03:00 Vladimir Ozerov <voze...@gridgain.com>: > Valya, > > Yes, in this design we will copy data into separate buffer on read. But > what is important - it will happen only for message which is split between > buffers. > > On Fri, Oct 14, 2016 at 2:33 AM, Valentin Kulichenko < > valentin.kuliche...@gmail.com> wrote: > > > Vladimir, > > > > We don't write length because be don't know it in advance. Sounds like > > you're proposing to marshal the message first and then copy it to the > write > > buffer. But that's actually our previous implementation and the whole > > purpose of direct serialization was to avoid this copy. > > > > The optimization for writes sounds interesting, though. > > > > -Val > > > > On Thu, Oct 13, 2016 at 3:51 AM, Vladimir Ozerov <voze...@gridgain.com> > > wrote: > > > > > Writes can be optimized even further: > > > 1) Write to *ByteBuffer *as long as there is a place in it. > > > 2) When it is full - invoke a callback which will submit it to the > > socket, > > > reset position to 0, and continue marshaling. > > > > > > This way we can probably get rid of write "state" at all. > > > > > > On Thu, Oct 13, 2016 at 1:17 PM, Vladimir Ozerov <voze...@gridgain.com > > > > > wrote: > > > > > > > Folks, > > > > > > > > I went through our so-called "direct serialization" and appears to be > > not > > > > very efficient to me. We never write message length. As a result we > > have > > > to > > > > constantly track what was written and what was not, and whether we > > have a > > > > room for the next write. The same goes for reader. As a result even > > > single > > > > "writeInt" is surrounded with multiple checks and writes. > > > > > > > > It looks like we can make our algorithm much more simple, > > straightforward > > > > and efficient if we add two things to every message: > > > > - Message length > > > > - Flag indicating whether it was written fully or not. > > > > > > > > If message was written fully to the buffer, we do no need to perform > > any > > > > checks during deserialization. To read int it is enough to call > > > > *ByteBuffer.getInt()*. To read byte array it is enough to call > > > > *ByteBuffer.getByte()*, etc. Simple and fast. > > > > > > > > And only if message was split into pieces on either send or receive > > > sides, > > > > which should not happen often, we may want to fallback to current > > > > implementation. Or may be we may copy such message to a separate > > buffer > > > > and still read them without any boundaries checks and > > "incrementStates". > > > > > > > > Thoughts? > > > > > > > > Vladimir. > > > > > > > > > >
