Bug in Bob Supnik's Emulator!

Tim Shoppa shoppa at alph02.triumf.ca
Sun Apr 5 06:43:25 AEST 1998


> > with gcc and running under Linux) is about twice as fast as a real
> > 11/73 for most CPU-intensive operations.  Speeds for I/O based
> > operations can range from incredibly faster to incredibly slower
> 
> 	Ok - I finally got around to retrying Bob's emulator.  This is using
> 	gcc 2.8.1 under BSD/OS 3.1 with a PPro 200 and the Dhrystone 2.1 (C
> 	language version) program.
> 
> 	Running under the emulator I get 555 dhrystones/second.  On a real
> 	11/73 I see 664 dhrystones/sec.

I suspect that the emulator will be quite slow on any math-heavy
benchmark - and your observations confirm this.  Doesn't Bob's
emulator do the FP operations by converting everything to IEEE
and back for each and every operand?

> > than a real -11, of course, and a lot of the interrupt and device
> > priority schemes seem seriously out of whack with how a real PDP-11
> 
> 	The line frequency clock seems to be acting strange.    When running
> 	the dhrystone program I see:
> 
> Measured time too small to obtain meaningful results
> Please increase number of runs
> 
> 	EVEN THOUGH the (wall clock) run time for 20000 dhrystones was 36 
> 	seconds.

On my cow-oreker's Pentium Pro, the line-time clock under Bob's emulator
appears to work fine, but it "misses" a lot of ticks when running on
my 7-year-old Alpha.  I've never looked at the logic to figure out exactly
what is going on, but I suspect that I couldn't emulate the interrupt/
priority structure any better than Bob's already done!

> 	Other benchmarks of possible interest:
> 
> 	A recompile of the 2.11BSD C compiler:
> 
> 11/44	9min 20sec
> 11/73   9min 33sec
> 11/93   6min 43sec
> emulated PDP-11 5min 25sec

For most "real" PDP-11 emulation uses this is probably a more realistic
benchark than the Dhrystone.  I know lots of currently-being-used-and-
maintained PDP-11 applications, and none of them are heavy on FP - all
the FP-specific stuff got migrated to a faster machine the instant
the faster machine became available.  (You'd be amazed at the awful
machines that I've seen people use *just* because it did their integral
faster.  Farms of I860's and I960's were the rage a couple of years ago,
and boy was that an icky development platform.)

> (BUT the 'time' reported with "time make" was 10min
> 			    4 sec)

The line-time-clock on Bob's emulator doesn't necessarily have anything
to do with reality.  On my cow-orker's 200 MHz pentium Pro, it ticks
about twice as fast as real time, but on my Alpha it'll often not tick
at all if there's something else keeping the (emulated) CPU busy.  I
think other emulators (like John Wilson's) put more emphasis on real-time
applications and probably emulate the line-time-clock more faithfully.

> 	Interesting that the emulated one is faster on this test even though
> 	the dhrystone rating is about 20% slower.

Again, I think the C recompile is probably a better benchmark - unless
someone's specifically interested primarily in FP emulation, which I think
is likely to be the exception.

Tim. (shoppa at triumf.ca)

Received: (from major at localhost)
	by minnie.cs.adfa.oz.au (8.8.5/8.8.5) id JAA23510
	for pups-liszt; Sun, 5 Apr 1998 09:29:27 +1000 (EST)
X-Authentication-Warning: minnie.cs.adfa.oz.au: major set sender to owner-pups at minnie.cs.adfa.oz.au using -f


More information about the TUHS mailing list