From owner-aic7xxx Fri Aug 14 16:41:56 1998 Return-Path: Received: (from majordom@localhost) by hub.freebsd.org (8.8.8/8.8.8) id QAA16240 for aic7xxx-outgoing; Fri, 14 Aug 1998 16:41:56 -0700 (PDT) (envelope-from owner-aic7xxx@FreeBSD.ORG) Received: from eta.ghs.com (eta.ghs.com [208.8.104.2]) by hub.freebsd.org (8.8.8/8.8.8) with ESMTP id QAA16221 for ; Fri, 14 Aug 1998 16:41:53 -0700 (PDT) (envelope-from ross@teraflop.com) Received: from random.teraflop.com (random.teraflop.com [192.67.158.207]) by eta.ghs.com (8.8.5/8.7.3) with ESMTP id QAA14404 for ; Fri, 14 Aug 1998 16:41:24 -0700 (PDT) Received: (from ross@localhost) by random.teraflop.com (8.8.8/8.8.8) id QAA21835 for aic7xxx@FreeBSD.ORG; Fri, 14 Aug 1998 16:41:24 -0700 (PDT) Date: Fri, 14 Aug 1998 16:41:24 -0700 (PDT) From: Ross Harvey Message-Id: <199808142341.QAA21835@random.teraflop.com> To: aic7xxx@FreeBSD.ORG Subject: Re: AHA2790UW has speed-limit problems ? Sender: owner-aic7xxx@FreeBSD.ORG Precedence: bulk X-Loop: FreeBSD.org > From: "Justin T. Gibbs" > > >> Shorten the cables. Ensure that the distance between all connectors on > >> the cable are equal. Use a forced perfect terminator. > > > >Shortening the cables might help, but what is the principle behind equalizing > >the inter-tap distance? > > Assuming that the drives on the chain offer similar capacitive loads, you > want the cable distance to be equal between ports to maintain a constant > impedance throughout the length of the bus. Yes, but only if the rise time of the incident wave is large relative to the transit time of the segment. Otherwise, the damage is done independently at each reflection point regardless of any symmetry. For a correctly matched purely resistive terminator, I would agree that the equally spaced loads provide an additional advantage of matching the energy output of the driver with the final termination, however, with a more advanced terminator this is dynamically adjusted, and with a typically mismatched resistive terminator things could get better, or they could get worse. > >(2) segments less than 0.3m (stub clustering lumps loads > > and aggregates the impedance mismatch); > > ... > > >I suspect that people think equalizing segment lengths helps because it forces > >them to use the recommended minimum stub (load) spacing. But what makes this > >win is the elimination of stub clusters, not the equalization of segment runs. > > This is a common misconception. The requirement, although dictated as an > explicit number by the SCSI spec, is really trying to say, the ratio of > stub length to stub spacing should be at least 1:3. I have a fairly > in-depth testing report provided by Adaptec to OEMs using their parts on > MBs, where they show with signal analysis that keeping the ratio is what > matters. Considering that the stub length of most drives is ~1" of PCB > trace, the inter-stub spacing can be as low as ~.1m. > -- > Justin It may be a misconception for many people, but in my case the intention is understood. Reading app notes is fine, but I've also done hundreds of spice simulations of transmission line designs. The real parameters are the rise time of the incident wave, the stub length, the segment length, and of course all the properties of the media and stub media. The ratio of rise time to stub length is really the most critical, but certainly the minimum segment will vary as an involved function of all these things. The problem is complicated considerably by the non-intuitive fact that your objective is _not_ to minimize the total signal distortion, but rather you want to minimize the magnitude of the largest single distortion at any sample point, even if this means increasing distortion that would be below the noise margin. That's why you don't want reflections to superimpose, even at the price of more noise. I suppose Matt Jacob basically got it right: we shouldn't be having this conversation at all. People should _not_ try to apply single-ended SCSI this way. High end drives have historically had the same cost in differential, and now we have FC and LVD alternatives which also solve this problem. The most price sensitive config is the point-to-point case, and for this SE SCSI is fine. If more people understood this and didn't insist on deploying SE SCSI so widely, then we would have more choices of vastly more reliable signalling technologies. So, there really isn't an excuse...just don't build systems like that one! --Ross Harvey To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe aic7xxx" in the body of the message