@amirm When you say „The noise from the CPU and operating system activity is so much more significant than any switch upstream.“ you mean that the NAS running the ROON core for example is having a far greater impact?
Generally the audio player is a user-space application. This means that packets are received by hardware and put in a DMA queue. The operating system then, after some random delay, transfers the packet to the kernel memory. The packet then goes through a bunch of kernel network code (more random delay) and ends up in a memory queue to be transferred (usually copied) to an application. The application gets some signal from the kernel that there's memory to read (again more delay) and it copies the memory out of the kernel to user space (again more delay). The application can then decode the packet and process it (again more delay). Application processing, of course, is preemptive multiprocessing in user space, so there's more random delay. Then what does the user-space code do? It sends some other memory block back down into the kernel to be DMA transferred to a sound card, which again has at least 1 memory copy and 1 DMA (or other mechanism) to the hardware (could be USB transfer to DAC etc).
All of these delays are in micro-seconds, maybe even 10s or 100s of micro-seconds, from the time the last bit of a Ethernet frame is received in hardware to the time user-space code has turned around and delivered audio to the sound card.
If you are playing music via Roon from a NAS, remember that the NAS has its own operating system and random delays reading data and packetizing it to send to Roon, which then has the above process on the core and then another network transfer to the Roon bridge where there's that same process yet again.
Random timing between Ethernet bits just shouldn't come into play. The randomization between whole Ethernet frames is huge compared to that, and the random delays in kernel and user space apps is huger.
yes, there are some ways do to direct NIC to user space app transfers and cut out most of the kernel, but I don't think streamers and audio players are using those DPDK-like techniques.