Date: Sun, 17 Feb 2002 17:20:57 -0800 From: Terry Lambert <tlambert2@mindspring.com> To: David Greenman <dg@root.com> Cc: Roy Sigurd Karlsbakk <roy@karlsbakk.net>, Dag-Erling Smorgrav <des@ofug.org>, Thomas Hurst <tom.hurst@clara.net>, hiten@uk.FreeBSD.org, hackers@FreeBSD.ORG, freebsd-questions@FreeBSD.ORG Subject: Re: in-kernel HTTP Server for FreeBSD? Message-ID: <3C7056F9.A9F37535@mindspring.com> References: <Pine.LNX.4.30.0202171935060.6486-100000@mustard.heime.net> <3C703A92.2EBD3E67@mindspring.com> <20020217170929.D80718@nexus.root.com>
next in thread | previous in thread | raw e-mail | index | archive | help
David Greenman wrote: > >"Zero copy" usually means "zero unnecessary copies"; but > >what someone thinks of as "necessary" is really based on > >their bias towards an existing implementation. > > "zero copy" these days has come to mean no copies that involve > the CPU. In my experiance, raw memory bandwidth to DMA packets > to/from main memory is not the bottleneck on modern hardware. I'll agree with your experience. At this point, the limiting factor is PCI bandwith, at least for general purpose hardware. The "AGP approach", where the main memory *is* the NIC memory, gets around this limitation, but doesn't kick the CPU out of the picture, even then. In practice, I've found that it is sometimes better to let the CPU do the copying, rather than the DMA engine on the device, since it's often faster at it. This assumes that the CPU is not otherwise engaged (e.g. doing cryptography), and there's enough spare cycles around the CPU can do the job faster. Moore's law and Intel really screw with the business model required to build special purpose network processors, don't they? 8-) 8-). -- Terry To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3C7056F9.A9F37535>