From mboxrd@z Thu Jan 1 00:00:00 1970 From: ebiederm@xmission.com (Eric W. Biederman) Subject: Re: Network virtualization/isolation Date: Sat, 25 Nov 2006 16:16:06 -0700 Message-ID: References: <78C9135A3D2ECE4B8162EBDCE82CAD77F0F698@nekter> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: , "Daniel Lezcano" , "Dmitry Mishin" , "Stephen Hemminger" , , "Linux Containers" Return-path: Received: from ebiederm.dsl.xmission.com ([166.70.28.69]:16521 "EHLO ebiederm.dsl.xmission.com") by vger.kernel.org with ESMTP id S935098AbWKYXRm (ORCPT ); Sat, 25 Nov 2006 18:17:42 -0500 To: "Leonid Grossman" In-Reply-To: <78C9135A3D2ECE4B8162EBDCE82CAD77F0F698@nekter> (Leonid Grossman's message of "Sat, 25 Nov 2006 17:17:03 -0500") Sender: netdev-owner@vger.kernel.org List-Id: netdev.vger.kernel.org "Leonid Grossman" writes: > I did not mean kernel bypass, just L2 hw channels that for > all practical purposes act as separate NICs - > different MAC addresses, no blocking, independent reset, etc. Yes. Nearly all of what you need for safe kernel bypass. >> In the worst case I might need someone to go as far as the >> Grand Unified Lookup to remove all of the overheads. Except >> for distributing the work load more evenly across the machine >> with separate interrupts and the like I see no need for >> separate hardware channels to make things go fast for my needs. >> >> Despite the title of this thread there is no virtualization >> or emulation of the hardware involved. Just enhancements to >> the existing hardware abstractions. > > Right, I was just trying to say that IOV support (likely, from multiple > vendors since > virtualization is expected to be widely used) would provide an option to > export multiple > independent L2 interfaces from a single NIC - even if only a subset of > IOV functionality would be used in this case. Agreed, and I think I understood that. My basic point was that it doesn't look to me like I need the hardware support, just that I can use it when it is there. The core advantage I see of the multiple queues, is in being able to split the processing of network traffic and interrupts among multiple cores. Eric