Use zero-copy (aka pipe/splice) for skbuff to enhance socketpair performance?

From: Fabian Franz
Date: Sun Sep 24 2006 - 16:45:03 EST


as far as I know pipe() is now much faster than socketpair(), because pipe()
uses the zero-copy mechanism.

However there are numerous applications using socketpair() instead of two
pipe()s, because they need exactly _one_ fd for read/write instead of 2.

Over 180 open unix domain sockets are on my system right now.

Many of this applications could gain a speedup if only they could change to a
different system call like epipe().

epipe() would behave like socketpair(), but internally use two pipes (one for
each direction).

It seems so easy to do, so I'm really wondering, why it hasn't been done, yet?

However new system calls always have the problem of portability. Though in
this case an epipe -> socketpair fallback would work.

So perhaps a better idea would be to enhance skbuff to use the new zero-copy
mechanism ... ?

The reason why I had the idea is that I had huge socketpair performance
problems, but with an old 2.4.27 kernel and a custom kernel module:

Almost 1 sec of latency!

Changing the fds provided from userspace to pipe()s did result in good

Might be a bug in my module though ...



PS: Please CC me as I'm not subscribed.

*** Consulting - Training - Workshops - Troubleshooting ***
@@@ LiveCDs (Knoppix), Debian, Remote Desktop Access (FreeNX) @@@

--- Fabian Franz --- --- consulting@xxxxxxxxxxxxxxx

Attachment: pgp00000.pgp
Description: PGP signature