Re: zero-copy TCP fileserving

Johannes Erdfelt (jerdfelt@sventech.com)
Thu, 3 Jun 1999 00:13:33 -0400

On Wed, Jun 02, 1999, Oliver Xymoron <oxymoron@waste.org> wrote:
> On Wed, 2 Jun 1999, Bjorn Wesen wrote:
>
> > Speaking of shortcuts, what is the status of some sort of direct
> > page -> skb TCP sending ?
> >
> > What I mean is, when tcp does a sendmsg, it shouldn't need to copy the
> > actual data into the skbuf, but merely the headers, and remember a ptr to
> > the data area (which would have to be locked down until the fragment is
> > acked or the connection breaks, of course). Then the network driver could
> > (if it supports s/g DMA) send the headers + data directly from memory.
>
> This simple zero-copy isn't really a win on most PC hardware as it takes
> about the same amount of time to do a checksum as it does to do a checksum
> and copy.

What about the new network cards which can do checksums in hardware?
Also, the decreased memory usage is also a (not as significant) win.

Albeit, the architecture to support hardware checksums isn't in place,
but that can be worked on.

> A more powerful scheme is being investigated by sct and a couple others. A
> related scheme worked out on FreeBSD and presented at a recent Usenix
> could cache checksums, for example.

Hmm, caching checksums? I wonder how that would work. It would only work
on data which is somewhat repeated. This would probably be a no win for
encrypted data. Albeit, I haven't read the paper yet, so I dunno how it
works.

JE

-
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/