Beg to differ, and differ big time.
You want to avoid the kernel for _Latency_ reasons. Bandwidth isnt the
issue since a syscall doing DMA to user space is the sme data copy rate.
Its when you get tons of small messages you want user mode hooks - things
like U-Net.
> On a PCI-SCI clustering adapter I used to work on in a previous (work)
> life, you needed to have ~1-2Kbyte DMA blocks before the DMA engine
> was worth using; otherwise just moving the data with the CPU was far
> superior.
You are priming the cache. I can well believe it. In networking I saw
bizzare artifacts like that - like where a 1500 byte packet copy into
an sk_buff if you avoided it and went direct to user space was no
different in cost.
Alan
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/