Date: Wed, 28 Jan 1998 14:06:46 -0800 From: Anne Hutton <hutton@ISI.EDU> To: Chris Csanady <ccsanady@friley585.res.iastate.edu> Cc: freebsd-hackers@FreeBSD.ORG Subject: Re: Gigabit ethernet cards for FreeBSD? Message-ID: <199801282206.OAA08022@tnt.isi.edu> In-Reply-To: Your message of Tue, 27 Jan 1998 19:52:19 -0600. <199801280152.TAA07604@friley585.res.iastate.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
> > >> > >> >Just wondering if there are any for FreeBSD , how they perform and how > >> >much they cost? > >> > >> Packet Engines makes one, as well as a full duplex repeater. The NIC's > >> are somewhere in the $1000-1500 range I believe. As for performance, it's > >> not quite gigabit yet. With 2 PPro200's, the TCP performance for larger > >> write sizes is still limited to around 20MB/s. > > > >we also found that TCP performance for the Myrinet (also a gigabit technology) > > >on PPro200's running FreeBSD was not good - around 160Mbps. > >UDP performance was better at around 300Mbps. However, to achieve higher > >throughput for host based IP forwarding we developed a driver capable of > "peer >DMA". This increased our throughput to 440Mbps. No changes were made to > the OS > >or protocol stack to achieve this. Ted Faber provided the references in an > >earlier email. actually I got the throughput wrong - it's 480Mbps for UDP. > > How are you doing the "peer dma?" Peer DMA involves the direct transfer of packets from the incoming network interface card (NIC) to the outgoing NIC, rather than staging packets through host memory. The resulting forwarding requires half the bus bandwidth. It should be possible on other NIC types, provided they have buffers and programmable processors. > Is it zero copy? Although it is > possible to get good performance without modifying the stack, this is > unfair, imho. There is very little hardware that is as smart as the myrinet > boards. They are definately nice.. > So we didn't modify the stack or use another API bypass of the stack but modified the driver. The IP header alone is copied into the hosts' main memory and the packet's payload remains resident on the incoming NIC. The techique is useful for host based IP forwarding...ie not useful in an end system. > > > >BTW what is the MTU on Gigabit ethernet - 1500? > > Yes, disappointing isn't it? This alone makes it impossible to do anything > reasonably efficient. our measurements would indicate the same. > Now it will most likely be a mediocre technology > for high performance networking. It will still be good for backbones though, > where individual machines can't push that much data. > > Is this not reason enough to improve the stack though? many changes have been put forward on how to improve the stack, data copying, servicing interrupts etc. Zero copy TCP has been implemented for solaris ( I think). Generally, (it seems to me) the way forward is still up for grabs. Our measurements also showed that the periperal bus was a bottleneck. Anne. -- Anne Hutton USC/ISI 4676 Admiralty Way, Marina Del Rey, CA 90292 Tel: 310-822-1511 ext. 211
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199801282206.OAA08022>