[Mono-list] NetworkStream.Read latency

[Atsushi Eno]

I wonder what if you use sgen-gc instead of the default one (i.e. configure
--with-gc=sgen). It should be used only for testing latency though.

Atsushi Eno

> Hi,
>
> I've run into an odd problem in a streaming TV application that I'm 
> building. I have a client and server on the same machine with the server 
> sending data as fast as Socket.Send() allows. The client uses 
> NetworkStream.Read() to retrieve data on-demand.
>
> Occasionally the client will underrun a quite sizable buffer set aside 
> to smooth out gaps in network delivery. Here's a sample from when that 
> occurs:
>
> Read 2048 bytes took 00:00:00.0000810
> Read 2048 bytes took 00:00:00.0000570
> Read 2048 bytes took 00:00:00.0000650
> Read 2048 bytes took 00:00:00.0000630
> Read 2048 bytes took 00:00:00.0000550
> Read 2048 bytes took 00:00:00.0000650
> Read 2048 bytes took 00:00:00.0000620
> Read 2048 bytes took 00:00:00.0990950
> Read 2048 bytes took 00:00:00.0001190
> Read 2048 bytes took 00:00:00.0000670
> Read 2048 bytes took 00:00:00.0996880
> Read 2048 bytes took 00:00:00.0001030
> Read 2048 bytes took 00:00:00.0000700
>
> There are small expected fluctuations in recv() latency across all of 
> the reads, but notice the two at suspiciously close to 0.1s. I added 
> timings probes to mono/metadata/socket-io.c to measure the 
> Socket_Receive_internal call and it showed no significant correlating 
> spike in latency, so I must assume this spike is being introduced 
> elsewhere in the Mono runtime.
>
> Might there be an obvious cause of this? The spikes are always so close 
> to 0.1s that it cannot be coincidental.
>