interesting paper at usenix...

[Perry E. Metzger]

There's a paper on TCP receive path being presented today. Dunno how I
feel about the ideas described but they are claiming very large
speedups in Linux. It might be worth looking at whether the ideas are
good and can be applied to NetBSD's stack.

http://www.usenix.org/events/usenix08/tech/full_papers/menon/menon_html/index.html

[Matt Fleming]

Post by Perry E. Metzger
There's a paper on TCP receive path being presented today. Dunno how I
feel about the ideas described but they are claiming very large
speedups in Linux. It might be worth looking at whether the ideas are
good and can be applied to NetBSD's stack.
http://www.usenix.org/events/usenix08/tech/full_papers/menon/menon_html/index.html

It is an interesting paper, though I'm a little skeptical that they
achieve the speed up they claim.

I don't think packet aggregation is feasible for NetBSD in general
because of this statement,

"Packet aggregation is done only for valid TCP packets, i.e., those with
a valid TCP and IP checksum. We verify only the IP checksum field of the
network TCP packet before it is used for aggregation. For the TCP
checksum, we assume the common case that the NIC supports checksum
offloading, and has validated the TCP checksum. This is because
verifying the TCP checksum in software would make the aggregation
expensive."

Matt

--
Posted automagically by a mail2news gateway at muc.de e.V.
Please direct questions, flames, donations, etc. to news-***@muc.de

[Matt Fleming]

Post by Matt Fleming
I don't think packet aggregation is feasible for NetBSD in general
because of this statement,
"Packet aggregation is done only for valid TCP packets, i.e., those with
a valid TCP and IP checksum. We verify only the IP checksum field of the
network TCP packet before it is used for aggregation. For the TCP
checksum, we assume the common case that the NIC supports checksum
offloading, and has validated the TCP checksum. This is because
verifying the TCP checksum in software would make the aggregation
expensive."

D'oh, I don't know why I thought it would need to be implemented across
the board, rather than as a compile-time option. Clearly not enough coffee
this morning.

I think the approach is worth looking into.

--
Posted automagically by a mail2news gateway at muc.de e.V.
Please direct questions, flames, donations, etc. to news-***@muc.de

[Jason Thorpe]

Post by Matt Fleming
I don't think packet aggregation is feasible for NetBSD in general
because of this statement,

Uh, it's really easy to determine that at run-time.

Post by Matt Fleming
"Packet aggregation is done only for valid TCP packets, i.e., those with
a valid TCP and IP checksum. We verify only the IP checksum field of the
network TCP packet before it is used for aggregation. For the TCP
checksum, we assume the common case that the NIC supports checksum
offloading, and has validated the TCP checksum. This is because
verifying the TCP checksum in software would make the aggregation
expensive."
Matt

-- thorpej


--
Posted automagically by a mail2news gateway at muc.de e.V.
Please direct questions, flames, donations, etc. to news-***@muc.de