Date: Fri, 11 Jul 1997 11:02:17 -0400 From: dennis <dennis@etinc.com> To: dg@root.com Cc: isp@freebsd.org Subject: Re: T1/T3 Upgrade Options? Message-ID: <3.0.32.19970711110214.00c9e520@etinc.com>
next in thread | raw e-mail | index | archive | help
At 11:17 PM 7/10/97 -0700, you wrote: >>A major factor to consider is that its very difficult to do 86Mbs (T3 is >>full duplex) >>with an addon card on the PCI bus because the sustained throughput rate >>is often pretty low. Other bus masters (ethernets, for example) will futhur >>reduce the burst capability. While PCI bursts to 128MB/s, very few PC >>products have sustained throughput rates over 100Mb/s. If you have a >>100Mbs ethernet card on the same bus (you'd have to) the number is >>cut in half. plus bus masters can't be controlled so you have contention >>issues. > > You're mixing bits and bytes, Dennis. The PCI bus is capable of >132 MBytes/sec which is 1.056 Gbits/secs. There are no problems at all >with the PCI bus being fast enough to handle 2 fast ethernet cards and >a few full duplex T3 interfaces. In fact most memory subsystems wouldn't >be limiting factor, either. There are problems with having enough CPU >power to queue small packets at full rate, however. Just one fast >ethernet is capable of about 140,000 packets per second, and you can't >even get a third of that rate with a PPro/200. So unless your average >packet sizes are large, you won't be able to keep the wire saturated. Oh, yes, 132, not 128, but I got the Bytes right :-) What your are confusing is "sustained throughput" and "burst rate". Bursts are meaningless, the sustained rate (which you need for a sync interface to avoid transmitter underruns) is the main issue. Do some RAM tests on PCI video and see what you get. Its highly MB dependent as well. Also, note that "fast ethernet cards" are custom ASICs which are capable of single-cycle bursts (required for 132MB/s), which are not nearly possible when using external logic and external devices for the communications (as SDLs board does). Once you start adding wait states, the bus becomes MUCH slower, and with single-cycle burst 100Mb/s ethernet cards stealing the bus whenever it wants, the T3 card (which should be the priority device) becomes secondary. With external RAM on the card (which the SDL card has), the PCI interface becomes much slower, and without it the card has to have guaranteed bus access and a large fifo to keep the transmitter filled. Basically, what I'm saying is that without a single device that handles the bus interface, includes RAM and does the HDLC communications, you cant get ANYWHERE NEAR the full PCI capabiltiy....and I doubt you could get satisfactory results. The math is real easy if you just take the burst rate and divide it up...unfortunately in practice is doesnt work nearly that efficiently. Dennis > >-DG > >David Greenman >Core-team/Principal Architect, The FreeBSD Project > >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3.0.32.19970711110214.00c9e520>