PDP-10.klh10/doc/dfkfb.txt

/* DFKFB.TXT - README for DFKFB KL timing diagnostic
*/
/* $Id: dfkfb.txt,v 2.3 2001/11/10 21:24:21 klh Exp $
*/
/*  Copyright © 2001 Kenneth L. Harrenstien
**  All Rights Reserved
**
**  This file is part of the KLH10 Distribution.  Use, modification, and
**  re-distribution is permitted subject to the terms in the file
**  named "LICENSE", which contains the full text of the legal notices
**  and should always accompany this Distribution.
*/

NOTE:
	DFKDB is not KLH10 software; it is a KL10 instruction timing
	diagnostic written by DEC and may only be used with a valid TOPS-10 or
	TOPS-20 license.  If present in the distribution, it is included
	purely as a convenience for developers.

People with a weakness for instant gratification, which is most of us,
will like DFKFB.  As soon as you have built a KN10-KL, you can
immediately run DFKFB without spending any time on installation or
configuration, and get an idea of how it compares with a real KL.

[1] Build a KN10-KL for your native platform.
	$ cd <distrib>/tmp/bld-kl
	$ make

[2] Go to the DFKFB directory, if present, and run it.
	$ cd ../../run/dfkfb
	$ ../../tmp/bld-kl/kn10-kl dfkfb.ini

[3] Compare output with other configurations, platforms, etc.
    Strut or sulk as appropriate.

As with all benchmarks, the DFKFB output must be taken with large
grains of salt.  Its results will be strongly influenced by the cache
behavior of your platform; with modern hardware it will usually run
entirely in the cache and thus will represent tne best speed you can
get.  When running a PDP-10 OS with real applications, you will get
many more cache misses and the following factors become more
important:

       - PDP-10 instruction mix.  Floating point is slow.
       - I/O response (fast physical disks make a difference).
       - Non-KLH10 processes competing for the native platform's CPU.

Still, it's fun to see how many multiples of a real KL you can run at.


THE REAL THING
==============

These numbers are from the 1978 EK-0KL20-IN-001 "KL10-Based
DECSYSTEM-20 Installation Manual", pages 10-27 and 10-28.

1 - BASIC CLOCK IS 40 NS.
2 - INDEXING TAKES 40 NS.
3 - INDIRECT TAKES 280 NS.
4 - INDEXING AND INDIRECT TAKES 320 NS.
5 - MOVEI TAKES 320 NS.
6 - MOVE FROM AC TAKES 440 NS.
7 - MOVE FROM MEMORY TAKES 480 NS.
8 - HRR FROM MEMORY TAKES 520 NS.
9 - STEOM 0 TAKES 560 NS.
10 - JRST TAKES 360 NS.
11 - JSR TAKES 680 NS.
12 - PUSHJ TAKES 840 NS.
13 - ADD FROM MEMORY TAKES 520 NS.
14 - MUL (9 ADD/SUB - 18 SHIFTS) TAKES 2.52 uS.
15 - DIV TAKES 5.58 uS.
16 - FIX A FLOATING POINT ONE TAKES 1.04 uS.
17 - FLTR AN INTERGER ONE TAKES 1.84 uS.
18 - FAD (1 RIGHT SHIFT) TAKES 1.88 uS.
19 - FAD (8 SHIFT RIGHT - 3 LEFT) TAKES 2.16 uS.
20 - FMP (7 ADD/SUB - 14 SHIFTS) TAKES 2.80 uS.
21 - FDV TAKES 5.72 uS.
22 - DMOVE FROM MEMORY TAKES 880 NS.
23 - DFAD (1 RIGHT SHIFT) TAKES 2.44 uS.
24 - DFAD (8 SHIFT RIGHT - 1 LEFT) TAKES 2.44 uS.
25 - DFMP (7 ADD/SUB - 32 SHIFTS) TAKES 4.92 uS.
26 - DFDV TAKES 10.32 uS.
27 - CONO PI TAKES 1.92 uS.
28 - CONI PI TAKES 3.36 uS.
29 - DATAO APR TAKES 1.56 uS.
30 - DATAI APR TAKES 1.76 uS.
31 - MOVE TO MEMORY TAKES 680 NS.
32 - LOGICAL SHIFT (35 PLACES LEFT) TAKES 640 NS.
33 - LOGICAL SHIFT (35 PLACES RIGHT) TAKES 760 NS.
34 - LOGICAL SHIFT COMBINED (71 PLACES LEFT) TAKES 1.12 uS.
35 - LOGICAL SHIFT COMBINED (71 PLACES RIGHT) TAKES 1.16 uS.
36 - INCREMENT BYTE POINTER TAKES 1.00 uS.
37 - INCREMENT AND LOAD BYTE TAKES 1.44 uS.
38 - INCREMENT AND DEPOSIT BYTE TAKES 1.80 uS.
39 - JFCL TAKES 880 NS.
40 - CAI TAKES 480 NS.
41 - JUMP TAKES 480 NS.
42 - CAM TAKES 600 NS.
43 - EQV AC TO AC TAKES 480 NS.
44 - EQV MEMORY TO AC TAKES 520 NS.
45 - SETOB TAKES 680 NS.
46 - AOS TO MEMORY TAKES 840 NS.
47 - EXCHANGE AN AC WITH AN AC TAKES 640 NS.
48 - EXCHANGE AN AC WITH MEMORY TAKES 840 NS.
49 - EXECUTE TAKES 640 NS.
50 - BLT MEMORY TO MEMORY TAKES 1.92 uS.
51 - BLT AC TO MEMORY TAKES 1.88 mS.
52 - DATAI TAKES 10.00 uS.
53 - DATAO TAKES 10.00 uS.

KLH's original text follows:

	Unfortunately, although I have output logs of DFKFB runs for a
variety of platforms, I do not have one for a real KL10 itself!
Hopefully someone will be able to contribute this from their archives.

The best information I can provide comes from a user-mode timing test
program I wrote called "ZOTZ"; the results I have are for a DEC-2065,
which is a KL10B with MCA25 cache.  Even though this is a user-mode
program rather than exec-mode as for DFKFB, these numbers should be
within a few percent of the real thing.  Copying this data into the
format of DFKFB gives the following simulated output, with "??"
inserted where nothing was available:

1 - BASIC CLOCK CYCLE IS ?? NSEC.
2 - INDEXING TAKES 36 NSEC.
3 - INDIRECT TAKES 269 NSEC.
4 - INDEXING AND INDIRECT TAKES 336 NSEC.
5 - MOVEI TAKES 269 NSEC.
6 - MOVE FROM AC TAKES 374 NSEC.
7 - MOVE FROM MEMORY TAKES 402 NSEC.
8 - HRR FROM MEMORY TAKES 443 NSEC.
9 - SETOM 0 TAKES 571 NSEC.
10 - JRST TAKES 303 NSEC.
11 - JSR TAKES 642 NSEC.
12 - PUSHJ TAKES 708 NSEC.
13 - ADD FROM MEMORY TAKES 438 NSEC.
14 - MUL (9 ADD/SUB - 18 SHIFTS) TAKES 2.42 USEC.
15 - DIV TAKES 4.70 USEC.
16 - FIX A FLOATING POINT ONE TAKES 874 NSEC.
17 - FLTR AN INTERGER ONE TAKES 881 NSEC.
18 - FAD (1 RIGHT SHIFT) TAKES 1.59 USEC.
19 - FAD (8 SHIFT RIGHT - 3 LEFT) TAKES 1.81 USEC.
20 - FMP (7 ADD/SUB - 14 SHIFTS) TAKES 2.56 USEC.
21 - FDV TAKES 4.82 USEC.
22 - DMOVE FROM MEMORY TAKES 608 NSEC.
23 - DFAD (1 RIGHT SHIFT) TAKES 2.06 USEC.
24 - DFAD (8 SHIFT RIGHT - 1 LEFT) TAKES 2.06 USEC.
25 - DFMP (7 ADD/SUB - 32 SHIFTS) TAKES 4.25 USEC.
26 - DFDV TAKES 8.71 USEC.
27 - CONO PI TAKES ?? NSEC.
28 - CONI PI TAKES ?? NSEC.
29 - DATAO APR TAKES ?? NSEC.
30 - DATAI APR TAKES ?? NSEC.
31 - MOVE TO MEMORY TAKES 573 NSEC.
32 - LOGICAL SHIFT (35 PLACES LEFT) TAKES 439 NSEC.
33 - LOGICAL SHIFT (35 PLACES RIGHT) TAKES 470 NSEC.
34 - LOGICAL SHIFT COMBINED (71 PLACES LEFT) TAKES 743 NSEC.
35 - LOGICAL SHIFT COMBINED (71 PLACES RIGHT) TAKES 745 NSEC.
36 - INCREMENT BYTE POINTER TAKES 924 NSEC.
37 - INCREMENT AND LOAD BYTE TAKES 1.21 USEC.
38 - INCREMENT AND DEPOSIT BYTE TAKES 1.60 USEC.
39 - JFCL TAKES 404 NSEC.
40 - CAI TAKES 272 NSEC.
41 - JUMP TAKES 342 NSEC.
42 - CAM TAKES 376 NSEC.
43 - EQV AC TO AC TAKES 402 NSEC.
44 - EQV MEMORY TO AC TAKES 742 NSEC.
45 - SETOB TAKES 571 NSEC.
46 - AOS TO MEMORY TAKES 539 NSEC.
47 - EXCHANGE AN AC WITH AN AC TAKES 545 NSEC.
48 - EXCHANGE AN AC WITH MEMORY TAKES 706 NSEC.
49 - EXECUTE TAKES 471 NSEC.
50 - BLT MEMORY TO MEMORY TAKES 1.58 USEC.
51 - BLT AC TO MEMORY TAKES 1.48 USEC.
52 - DATAI TAKES ?? NSEC.
53 - DATAO TAKES ?? NSEC.