> > When transferring one or more pages via a page-alligned
> > buffer and normal read() or write(), VM tricks will be
> > used to avoid copying the data. If you touch the page
>
> You mean mmap(). You can do that already but it isnt the win you might
> think. In fact for some operations mmap is slower just due to the mmu bashing
> overhead and TLB misses.
>
> Its non intuitive that mmu operations cost more than a bulk copy but its
> frequently the case. sendfile() works nicely because it is working on data
> that isnt user mapped. O_DIRECT raw I/O has overheads too and its not clear
> where the break even points are
Yup. That's true. Like it's also nonintuitive sometimes that PIO can even
beat, in some cases, DMA (some graphics fifo devices, e.g.).
>
> > They've gotten 960 megabits/sec out of a gigabit Ethernet card
> > with this. Not stable yet.
>
> I think Jes was getting this in Linux TCP/IP without zero copy.
What was the CPU utilization?
-matt
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sat Sep 23 2000 - 21:00:17 EST