| T.R | Title | User | Personal Name
 | Date | Lines | 
|---|
| 97.1 | It would require some hardware support | CSC32::B_HIBBERT | When in doubt, PANIC | Tue May 06 1997 12:33 | 9 | 
|  |     A PC history buffer would require some hardware support.  We did have 1
    product that had a PC history buffer implemented, the VAX 9000.  It had
    a 4K entry buffer that was written to the console storage on restarts
    and reboots.  This buffer was EXTREAMLY usefull in diagnosing some of
    the less obvious crashes.  Unfortunately, this functionality has not
    been added into any CPU since the 9000.  
    
    Brian Hibbert
    
 | 
| 97.2 | Yea it was nice but what about now not then? | NEASYS::MERCURE | Whats burning??? | Mon May 12 1997 08:32 | 16 | 
|  |     By hardware support, do you mean a redesign of the 21164 alpha cpu itself
    or the addition of hardware on an ALPHA CPU module? It's obvious the
    latter would be a lot easier to implement.  I have a customer (Citibank)
    with a very complex IS configuration. It is not unusual for them to
    log several UNIX core-dumps  per week. They believe they haven't been
    getting complete and comprehensive analysis on their dumps and have
    asked this question on more than one occasion.
    
    Since the PC History buffer is so useful, and the need for it has not
    deminished (if anything it has increased), I'll ask the ALPHA
    archetects. Why do we not employ a PC History buffer?  Please do not
    use the same old (and quite lame) cost to revenue ship excuse. Surely
    the Turbolaser has taught us that design and manufacturing cost savings
    are lost on the other end when we end up with a machine that is
    difficult to diagnose. 
    
 | 
| 97.3 | Things have moved on since 8600s | BBPBV1::WALLACE | john wallace @ bbp. +44 860 675093 | Tue May 13 1997 15:17 | 14 | 
|  |     You can't take an existing CPU chip/module and "just add hardware" to
    get a PC trace buffer. This applies in general to any CPU which has an
    instruction cache, and not just to Alpha(s). The signals you'd need
    just don't come off the chip. New features such as speculative
    execution just make this worse.
    
    What other possibilities might help in solving your unexplained
    crashes? Of course ideally they have to cost nothing to implement, as
    Digital Semi or the SBU pick up the implementation cost, but the lack
    of decent crash recovery tools doesn't show up as a cost on anyone's
    bottom line. Except Digital's, of course, and no-one's in charge there.
    
    regards
    john
 |