Re: GNAT.Sockets Streaming inefficiency

[masterglob@gmail.com]

Le vendredi 9 juin 2017 17:24:24 UTC+2, Dmitry A. Kazakov a écrit :
> On 2017-06-09 15:30, gautier_niouzes@hotmail.com wrote:
> 
> > It could be related to a performance bottleneck in 'Read and 'Write,
> > perhaps more than to the nature of the stream (socket vs. file vs. ...).
> 
> 1. Without NODELAY. Writes are buffered and spilled out either when the 
> buffer is full or when the coalescing algorithm timeout expired. With 
> string length is not a multiply of the buffer and moderate output rate 
> or synchronized output, the last frame carrying the string will be 
> always delayed. So the observed latency will be more or less the 
> coalescing algorithm timeout.
> 
> 2. With NODELAY. Individual writes are sent in independent frames. If 
> 'Output or 'Write write individual characters separately, plus the 
> string bounds, then the network gets flooded with short frames and the 
> latency depends on how good the stack, the OS, the network 
> infrastructure are.
> 
> In either scenario there is nothing good to expect.
> 
> -- 
> Regards,
> Dmitry A. Kazakov
> http://www.dmitry-kazakov.de

I currently do not use "NO_DELAY", but the thing is that the "producer" process is 100%. This tends to indicate that it is not this kind of issue (I would have long execution time, but not a 100% CPU usage).

I have to try other implementations proposed (like using something else than Stream 'write & 'Read)