Conversation

Replying to

looking into dpdk it seems that this could in principle be used for hard real time applications dedicated cores interact with nics directly using shared memory (dram or cache) and cores uses polling receive

Replying to

the absolute lowest latency answer is use a smart nic, the second lowest latency is run in a kernel thread, after that is probably cpu affinity and io_uring. though it also matters whether you're talking about worst-case, average, or 90th percentile

Replying to

and

for worst case you need to worry about the scheduler

Replying to

mostly interested in worst case in the order of 1ms. this is for sending real time audio

Replying to

and

intel e1000 seems to be supported in DPDK so basically if i do UDP or raw Ethernet this shouldn't be too hard to try out... the NIC doesn't need to do anything smart i think?

Replying to

I haven't used dpdk but I've heard only bad things.

@sargun

is a good person to ask

Replying to

and

Exactly the kind of thing I'm looking for before diving in. Bad in what sense?

Replying to

and

DPDK is basically a framework for writing a "unikernel" for packet processing. ...The unfortunate part is that you're effectively having to co-exist with Linux. That includes the scheduler, other tasks, etc...

Replying to

and

I'd actually strongly suggest XDP with a userspace ring buffer, instead of DPDK.

Replying to

and

Ok I see. Thanks! Makes sense then that DPDK is mentioned in the context of feedback control (e.g. PID regulator) where you would typically use a separate microcontroller or DSP with RTOS or bare mainloop

Replying to

and

another Mnemos use case :-)

2:24 PM · Aug 29, 2022Twitter Web App