Conference hydra::axp-developer

3430.0. "Orbtek Corporation" by HYDRA::BRYANT () Wed Apr 02 1997 17:39

    Company Name :  Orbtek Corporation
    Contact Name :  Chuck Broadus
    Phone        :  801/484-8777
    Fax          :  
    Email        :  
    Date/Time in :   2-APR-1997 16:37:36
    Entered by   :  Pat Bryant
    SPE center   :  MRO

    Category     :  NT
    OS Version   :  4.0
    System H/W   :  Alpha


    Brief Description of Problem:
    -----------------------------
	Pat;

	As we discussed, please take this call and help them determine the
 source of their alignment exceptions.  I suspect that there may be several
 places that they are getting the exceptions, but if you can find and fix
 to top several, you may gain a tremendous performance difference.  Please also
 let me know what they are referring to by (fp emulation)?  I would also
 like to know exactly what platform (alpha) they are running this on and 
 what the performance results are for

	system							time

	intel p6/200
	alpha (wm=what model) orig
	alpha (wm=what model, no exceptions)

	As a secondary issue, see if you can help them with the odbc driver
 question.

	Thanks/regards

	MikeG


From:	MRMTS::MRMTS::MRGATE::"DOHENY::DELREYA1::POIANI.MICHAEL" 31-MAR-1997
21:47:12.64
To:	MROA::MGREENFIELD
CC:	
Subj:	Porting Request

From:	NAME: MICHAEL POIANI@PHX <POIANI.MICHAEL@DELREYA1@DOHENY@IVO>
To:	NAME: MICHAEL R GREENFIELD <Michael Greenfield@MRO>
CC:	NAME: BILL ALLEN <BILL ALLEN@PHX>,
	NAME: MICHAEL POIANI <MICHAEL POIANI@PHX>,
	NAME: SUNIL KULKARNI <Sunil Kulkarni@MKO>

Hi Mike,

I received your voice mail today and I am following up with this E-mail you
requested.

The Customer is:

Chuck Broadus (engineering manager)
Orbtek Corporation
3030 S Mainstreet
Salt Lake City, Utah 84115
801-484-8777

Orbtek builds medical eye/retna scanning turnkey systems. The integrate
hardware with their application software which does calculations on the data
received from the eye scan hardware. Their application does many floating point
math calculations.

They built their application on WindowsNT on Intel. They currently embed a
Pentium Pro 200 MHz with each eye machine. We are trying to get them to embed
an Alpha Miata instead. They need as much performance as they can get,
customers are willing to pay more for performance. Intel does not have adequate
FP performance as they feel they need. The opportunity is to become Orbtek's
embedded platform choice estimated as high as 20 Machines per month.

Their application code was written with Visual C++ for Intel. Their application
interfaces to an Access database to store results and pull data. The Access
database was written on the Intel box, and is Accessed through the ODBC driver
on Intel. They literally take the .mdb access database file and copy it over to
the Intel box that runs the application. The application is able to access the
.mdb file through the ODBC driver included with Visual C++ for Intel.

I gave the customer a Miata workstation and loaded Visual C++ for RISC/ALPHA
4.0a. The customer moved their VisC++ project over to the Alpha and recompiled.
They had a couple of problems:

1. Could not get ODBC driver to load. We tried many ways to do this with the
VisC++ RISC CD and nothing seemed to work. I also found out that the ODBC
driver that comes with the kit may not be what we need. I was told we may need
an ODBC JET driver in order to access an Access database. This issue needs
resolving. I could not find sufficient expertise to close this out or verify my
information. Customer is willing to look at alternatives to Access if
necessary.

2. Customer could comment out Database access portions of code. He successfully
compiled and ran a portion of his application, the section with all the FP
calculations. The results were very poor. The section of code ran on a Pentium
Pro 200 in approx. 20 seconds. The same section ran on the Alpha in about 2
Minutes.

I had him run Perf Mon and he found that he was getting approx. 600,000
alignment fix-ups per second, so there are data alignment problems. I also
think they may be running Floating Points in emulation, from past experiences.

The customer does not have RISC porting experience. They only know the Intel
C++ environment. They are willing to work with us on getting this ported
cleanly. They are willing to start a new project and do this from scratch if
deemed necessary.

I need help getting this going in the right direction. I would like someone to
call the customer directly (use above name and number) and scope out the
situation, and work with Orbtek in the process of porting. I believe Orbtek may
be willing to send your folks the actual code if you think you can't work this
by phone, you may want to ask them.

If there is any additional information I can provide please call me, I am:

Mike Poiani
Workstation Tech Support
602-240-3565
[email protected]
mike poiani @phx
swam1::poiani_mi

Please feel free to go right at it and call the customer and get started. I
will maintain communications with the customer and they will keep me up to
date. I don't want to bottleneck this opportunity.

Sincerely,

Mike Poiani

Mike,

I had a conversation with Chuck Broadus today.  He thought that he had found all
his fp exceptions via the debugger and therefore could not understand why he had
so many faults.  I found out it was because he has the AXPALIGN env var set to
/ENABLE.  Instructed him how to set it to /DISABLE and rerun his app in the
debugger to find the rest.  Also explained to him why he is getting the faults
on AXP and not INTEL.  Told him I would provide him with a tool (NTFLOP) that is
useful for pinpointing floating point exceptions.  He told me in 20 days he goes
to a trade show for so he won't have much time to work on this.

ODBC--The driver he needs is on VC++4.2 however he is unable to install it. 
I've heard problems doing this so I'll research and call him back.

Pat Bryant

Its log as RAID 24961

I guess his son will be taking over this task.  The benchmark now is:
Intel :39 seconds / Alpha 2:00 minutes.

I showed him how to use Performance Monitor.  He is seeing 700,000 alignment
fixups per second.  He has NTFLOP.  I told him to run it to weed out where
exactly the alignments are occurring.

email address is [email protected]