        
        Ŀ
        ĳ
          ͻ  ĳ
                                                           ĳ
                     Steve's Indictment of the             ĳ
              Small Computer Systems Interface (SCSI)      ĳ
                                                           ĳ
                                or                         ĳ
                                                           ĳ
                "Why I wish I'd never bought SCSI"         ĳ
                               ....                        ĳ
                                                           ĳ
          ͼ  ĳ
        ĳ
        
        
        
        It's usually my PC successes that I choose to share, but about
        six months ago I made what I now consider to have been a big and
        expensive mistake when I purchased a SCSI hard disk drive for my
        main personal system. Today I'm six months older, quite
        exhausted, a bit poorer, much wiser ... and I'm stuck with a
        quite expensive SCSI drive that I'm still trying to justify.
        Sigh. This week and next I'll explain what happened, in the hope
        that you'll be prevented from following too closely in my
        footsteps.
        
        It all began when I decided to try flying right up to the edge
        of contemporary personal workstation performance. Today that
        means running a fast caching 486 on an EISA bus with the
        appropriate EISA bus peripherals. So I gave my favorite
        motherboard company, Mylex, a call in Fremont California to see
        what they were up to in the way of 486 motherboard technology.
        It turned out that they were just putting the finishing touches
        on a pair of lovely new 25 megahertz EISA 486's. The less
        expensive of the two lacked external processor RAM caching,
        relying exclusively upon the small 8k cache built into every 486
        chip. Mylex's high end 486 motherboard sports a fast 128k byte
        posted-write cache and delivers extremely high-end performance.
        
        Well, the non-caching 486 found itself an immediate home running
        Novell's Netware 386 LAN software where it has performed like an
        Olympic champion ever since. Of course, since I wanted to find
        out what ultra top-end killer performance really felt like, I
        took the caching 486 montherboard as my own. Up to that point I
        had been running with a 20 megahertz Mylex caching 386
        motherboard and Western Digital's very nice WD1007A-WA2 16-bit
        ISA ESDI controller with a pair of fast half high Miniscribe
        3150E 150 megabyte ESDI drives. This setup had been operated
        flawlessly forever.
        
        The 3180E ESDI drives each have 7 heads, 35 sectors per track,
        and 1250 cylinders. The high cylinder count would normally be a
        problem for a PC since the PC's venerable BIOS can only address
        1024 cylinders worth of drive data. But the Western Digital
        controller elegantly sidesteps this limitation by "translating"
        the 3150E's physical dimensions into something the PC's BIOS can
        readily handle. In the case of the 3180E, the WD controller
        translates the drive's dimensions so that it appears to have 16
        heads, 63 sectors per track, and a total of only 303 cylinders.
        
        The final blessing of this controller is it's faithful emulation
        of the classic Western Digital register-level chip set which
        allows the controller's to operate without a BIOS of its own,
        since it's then able to utilize the WD compatible BIOS which is
        part of any AT-class machine. This allows the controller's own
        high memory consuming BIOS ROM to be replaced with RAM memory
        provided by any 386 memory manager. Device drivers and TSR's can
        then be placed into this region. Overall this solution provides
        a stunningly compatible, high performance power-user solution
        which I'll heartily recommend day in and day out ... especially
        after getting burned when I was seduced by the dark side and
        strayed into the land of SCSI!
        
        I was captivated by Mylex's incredible 32-bit EISA bus, 4
        megabyte caching, bus mastering SCSI host adapter. Their model
        DCE376 controller is designed around Intel's controller-oriented
        version of the 386 and promised performance beyond anything else
        I'd ever experienced. As we'll see, a powerful synergy is
        created when a bus mastering controller is blessed with a big on
        board cache memory. Since I was giving up my prior pair of 150
        megabyte ESDI's, I went searching for a SCSI drive with
        comparable storage. I ended up with a big, fast, and expensive
        Maxtor 330 megabyte full height SCSI drive which I now wish I
        didn't own. My persistent unhappiness is no fault of Maxstor's
        however, since the drive has operated flawlessly from its first
        moment of use. As far as I can tell Maxtor has made a terrific
        drive. I just wish that I didn't now own a large and expensive
        SCSI drive.
        
        As nearly as I can tell SCSI is an inherently good thing. It was
        designed recently and has received the well-considered attention
        of many of our industry's most talented interface designers.
        People in the know feel that SCSI will own the future, but the
        uncertainty lies in determining when that future will finally
        happen. I can state with clear certainty that most SCSI
        controllers currently being sold in this industry need another
        generation of evolution in order to mate effectively with the
        incredibly kludgy heritage of the PC.
        
        Looking back I realize that my biggest problem was probably that
        I decided to upgrade many hardware and software components of my
        system all at once. I changed my motherboard, hard disk
        controller, and hard disk controller technology (from ESDI to
        SCSI) at the same time as I began working with Windows 3.0, a
        new unreleased fax board and both of Quarterdeck's and Qualitas'
        then-unreleased memory managers. When I got my new system all
        reconfigured and plugged together, nothing worked. Nothing. The
        whole expensive mess just sat there like a large humming power-
        hungry lump. Every so often a drive light would flicker briefly,
        as if this huge monster was slowly struggling toward
        consciousness. After a while I decided that it was simply
        teasing me, daring me to tackle the enigma that it had become.
        
        Naturally I did. I've grown so utterly dependent upon my
        computer system for my daily operating life that I could
        scarcely endure this state of affairs. So I stepped back to the
        earlier beta of 386 to the Max which I'd been using for months,
        and things got somewhat better. It seemed that this unreleased
        Max was in transition from 16 to 32 bit segments internally and
        was having some trouble with the scope of my "maxed out" 486. Of
        course this meant that I had to suspend my work with Enhanced
        mode Windows 3.0, but that was of lesser concern. Still, the
        entire system would completely freeze up from time to time, my
        trusty old Everex tape drive would lie there groaning, and I
        still couldn't fax. Wondering whether the problem might lie with
        the Mylex SCSI controller, I backed away from it too. But since
        I had just invested in this very lovely 330 megabyte Maxtor hard
        disk I went in search of SCSI alternatives.
        
        So I first switched over to Adaptec's new 1542 bus mastering
        SCSI host adapter. Near as I can tell this is a nice board, but
        it sure doesn't get along with QEMM 5.1! Whenever they are
        together in the same system things lock up immediately.
        Someone's got a problem there. One of the things I did notice
        immediately about the bus mastering Adaptec 1542 was that it was
        much slower than the Mylex. This got me started off on a
        separate adventure to understand what was really going on with
        SCSI system performance. I'd grown tired of heading all the
        claims about so-and-so's ultra high SCSI adapter performance and
        decided to find out for myself.
        
        So I assembled a complex series of SCSI performance benchmarks,
        to experiment with different styles, rates, and amounts of data
        transfer between CPU and hard disk subsystems. I ran ten quite
        different SCSI adapters through their paces. Some were bus
        masters, some had fifty chips and Motorola 68000 processors
        whereas some others had as few as three lonely IC's on a small
        PC board. Would you believe that only one single exception every
        SCSI adapter performed EXACTLY IDENTICALLY!
        
        The simple fact is that the physical transfer onto the hard disk
        drive's magnetic medium is the complete and total bottleneck in
        any SCSI system, and the entire issue of SCSI adapter
        "performance" is muted underneath the far greater issue of SCSI
        hard disk performance. One thing really did piss me off though.
        I happened to have a tiny little 3.5" half high 200 megabyte
        SCSI drive on loan from Conner Peripherals. This little
        screamingly fast and sexy gem consistently delivered 300% higher
        read performance than my oh-so-expensive and apparently brain
        dead Maxstor SCSI drive! Whatever the Conner drive is doing in
        the way of internal read-ahead caching sure works! Other SCSI
        drive makers would do well to pay attention.
        
        The single stand-out exception in SCSI adapter performance was
        delivered by Mylex's bus mastering controller with its 4
        megabyte on-board cache. The beauty of such a large on-board
        cache is that the drive's performance is then effectively
        decoupled from the CPU system. The adapter can accept a write
        request from the system and utilize the inherent performance
        edge provided by bus mastering to move the system's data into
        its cache for background writing to the hard disk.
        
        Okay, let's recap the salient points we've addressed up to this
        point:
        
        First, an ultra-compatible non-SCSI solution like, for instance,
        that provided by the WD1007 and several other EDSI or RLL
        controllers provides absolutely state-of-the-art performance
        equalling anything which SCSI can deliver.
        
        Second, with only two exceptions that I'm aware of, SCSI
        controllers all greedily consume a large chunk of precious high-
        memory. Their manufacturers' claim that this high memory BIOS
        allows the controllers to operate much faster, but I've
        determined conclusively that this is just so much utter hogwash.
        And quite frankly, I've grown sick to death of hearing it.
        There's nothing whatsoever inherently "lower performance" about
        emulating the standard WD chip set registers. The Mylex and DPT
        SCSI controllers that specifically do not bring along their own
        high BIOS, but instead emulate the WD register set, perform at
        the top of the SCSI performance charts. I'd MUCH RATHER leave my
        high-memory free for loading network drivers and TSR's. Wouldn't
        you?
        
        Third, the much touted bus-mastering feature of several "high
        performance" SCSI controllers is meaningful ONLY when the
        controller also contains a significant size on-board cache.
        Otherwise the controller's actual performance delivery is
        completely bottlenecked by the rate of physical data transfer to
        the hard disk medium. The two Mylex DCE376 and DC376 SCSI
        controllers were the only bus-masters to make any difference
        whatsoever due to their large on-board caches. Be very wary of
        bus-mastering claims too.
        
        Fourth, SCSI controllers seem to inherently beg compatibility
        problems. I never was able to get the Adaptec 1542 controller to
        operate with QEMM 5.1. What if you were stuck owning both
        products? Almost without exception all SCSI controllers come
        with large sets of "drivers" for various operating systems. This
        issue doesn't even arise with controllers which offer WD
        register level compatibility since compatibility with such
        controllers is inherently built into all operating systems.
        
        The final bit of venom I feel toward the largely over-blown
        claims for SCSI surrounds all the hogwash we're continually
        hearing about single controller to multi-device connectivity.

        The idea is that eight individually intelligent SCSI devices can
        all be bussed together onto a single SCSI bus. One of them,
        typically the last numbered device, number 7, is the so-called
        "host adapter." This is the controller which interfaces the SCSI
        bus to the "hosting" PC. This leaves seven other devices inter-
        connectable. The way the still-fictitious story goes, in
        addition to multiple hard disk drives, you might have an image
        scanner, a tape backup unit, a CD-ROM or two, and perhaps even a
        network interface all communicating simultaneously along a
        single high-performance SCSI bus. Yeah, sure, and Windows 1.0 is
        a great platform for multitasking on a 4.77 Mhz PC too!
        
        Nearly as I can tell, SCSI is currently suffering from a lack of
        standardization. Historically such industry acceptance follows
        adoption by some super huge giant. However in this case, IBM's
        recent infatuation with SCSI worries me. They'll probably find
        some way to shoot SCSI in the head rendering it brain dead and
        useless to all of us. But I hope it doesn't happen. Despite all
        my current misgivings, SCSI is good news ... but not yet ready
        for prime time.
        
        For now, you're really better off treating the promise of SCSI
        as just that ... an undelivered promise supported by plenty of
        hot air.
                             - THE END ... for now -


