shithub: 9ferno

ref: 8d0725d57a988ed0ed2ef0a33e0cebe48a85b5db

View raw version
.TL
Limbo profilers in Inferno
.AU
J R Firth
.AI
Vita Nuova
.br
13 June 2002
.SP 4
.NH 1
Introduction
.LP
Currently there are three application level profiling tools in the
Inferno package. 
.I Prof
is a time profiler which, by sampling, can provide statistics
on the percentage of time spent on each line of limbo source. 
.I Cprof
is a
coverage profiler which provides the execution profile for limbo source
code. It can accumulate results over a series of runs to allow full
coverage testing. Finally, 
.I mprof
is a memory profiler which provides
statistics on the amount of memory used by each line
of limbo source.
.LP
Two gui versions of these tools currently exist. 
.I Wm/cprof
shows the
coverage per module and highlights those lines which have not been
executed or have been only partially executed. 
.I Wm/mprof
shows the memory
usage per module and highlights lines with high memory allocation in
darker shades of red.
.I Prof
itself does not have a gui equivalent as it
was originally written to determine why acme was so slow when using it's
global editing command. A gui for it was not a requirement at that stage.
.LP
All these tools use a common library module
.CW /module/profile.m
and
.CW /appl/lib/profile.b
that acts as the direct interface with the kernel profiling device.
.LP
Note that none of these tools give kernel profile statistics. For that, the devmem driver should be used.
.LP
Although the use of these tools is very similar, there are a few differences
when it comes to interactive testing as the profilers were written to
answer different questions. Thus 
.I prof
tries to determine 'why is blah so
slow ?', 
.I cprof
tries to accumulate coverage records over time and 
.I mprof
tries to
give a series of memory statistics at intervals during the execution of
a program or series of programs.
.NH 1
Prof
.LP
The time profiler works by sampling. A kernel procedure sleeps for the
given sample time and then notes the particular dis instruction currently
being executed before repeating the process. After many such samples, an accurate profile can be obtained.
.LP
At it's simplest we can profile a particular command by giving the command
to execute followed by any arguments to the command eg
.sp
.RS
	prof wm/polyhedra
.RE
.sp
profiles the polyhedra displayer. After letting the latter run for a reasonable
amount of time, we exit and then get the following statistics.
.br
.DS
Module: Wmlib(/dis/lib/wmlib.dis)

34	0.06		str = load String String->PATH;

**** module sampling points 1 ****

.DE
.DS
Module: Bufio(/dis/lib/bufio.dis)

340	0.06			n := 0;
341	0.12			while(b.index < b.size){
342	0.99				(ch, i, nil) = sys->byte2char(b.buffer[0:b.size], b.index);

**** module sampling points 19 ****

Module: Polyfill(/dis/math/polyfill.dis)

37	10.80		for(i := 0; i < n; i++)
38	19.86			b0[i] = b1[i] = ∞;
57	0.06		p.y += y;
58	1.18		dst.line((left, y), (right, y), Endsquare, Endsquare, 0, src, p);
63	0.12		prevx := ∞;
64	9.93		for(x := left; x <= right; x++){
65	20.61			if(z+e < zbuf0[k] || (z-e <= zbuf1[k] && x != right && prevx != ∞)){
66	6.46				zbuf0[k] = z-e;
67	5.71				zbuf1[k] = z+e;
68	0.74				if(prevx == ∞)
69	0.74					prevx = x;
71	0.12			else if(prevx != ∞){
72	0.25				fillline(dst, prevx, x-1, y, src, p);
73	2.61				prevx = ∞;
75	4.35			z += dx;
76	3.17			k++;
78	0.06		if(prevx != ∞)
79	0.87			fillline(dst, prevx, right, y, src, p);
80	0.06	}
152	0.06			return (vx/z, mod);
186	0.06			sp.dzrem = mod(sp.num, sp.den) << fixshift;
187	0.06			sp.dz += sdiv(sp.dzrem, sp.den);
217	0.62			for(q = p = 0; p < ep; p++) {
218	0.37				sp = seg[p];
220	0.12					continue;
221	0.12				sp.z += sp.dz;
222	0.19				sp.zerr += sp.dzrem;
223	0.12				if(sp.zerr >= sp.den) {
224	0.19					sp.z++;
225	0.25					sp.zerr -= sp.den;
226	0.25					if(sp.zerr < 0 || sp.zerr >= sp.den)
227	0.25						sys->print("bad ratzerr1: %d den %d dzrem %d\n", sp.zerr, sp.den, sp.dzrem);
229	0.31				seg[q] = sp;
230	0.31				q++;
233	0.25			for(p = next; seg[p] != nil; p++) {
234	0.06				sp = seg[p];
247	0.12			ep = q;
248	0.06			next = p;
257	0.06				continue;
260	0.62			zsort(seg, ep);
262	0.25			for(p = 0; p < ep; p++) {
263	0.19				sp = seg[p];
264	0.06				cnt = 0;
265	0.06				x = sp.z;
266	0.25				ix = (x + onehalf) >> fixshift;
267	0.06				if(ix >= maxx)
271	0.06				cnt += sp.d;
272	0.12				p++;
273	0.25				sp = seg[p];
275	0.06					if(p == ep) {
277	0.06						return;
279	0.06					cnt += sp.d;
280	0.12					if((cnt&wind) == 0)
283	0.19					sp = seg[p];
286	0.25				ix2 = (x2 + onehalf) >> fixshift;
291	1.92				filllinez(dst, ix, ix2, iy, zv+ix*dx, er, dx, k+ix-zr.min.x, zbuf0, zbuf1, src, spt);
293	0.06			y += (1<<fixshift);
294	0.31			iy++;
295	0.06			k += xlen;
296	0.06			zv += dy;
298	0.06	}
310	0.06				done = 1;
311	0.12				q--;
312	0.25				for(p = 0; p < q; p++) {
313	0.87					if(seg[p].z > seg[p+1].z) {
367	0.06			t = a[0]; a[0] = a[i]; a[i] = t;
373	0.06				while(i < n && ycompare(a[i], a[0]) < 0);
379	0.12				t = a[i]; a[i] = a[j]; a[j] = t;
384	0.06				qsortycompare(a, j);

**** module sampling points 1584 ****

Module: Polyhedra(/dis/wm/polyhedra.dis)

327	0.12		return (int (geo.sx*v.x)+geo.tx, int (geo.sy*v.y)+geo.ty);
471	0.06					if(allf || dot(geo.view, newn[j]) < 0.0)
472	0.06						polyfilla(fv[j], new, newn[j], dot(geo.light, newn[j]), geo, concave, inc);
496	0.06			ap[j] = map(vtx, geo);
512	0.06			if(a <= -LIMIT || a >= LIMIT)
531	0.06			fillpoly(RDisp, ap, ~0, face, (0, 0), geo.zstate, dc, dx, dy);

**** module sampling points 7 ****


**** total sampling points 1611 ****
.DE
.br
The output lists all lines in all modules with a sampling point. Each line
shows the line number in the corresponding source file, the percentage of
time spent on that line and the source code. We can see that about 60% of
the sampling points occur on lines 37, 38, 64 and 65 of the Polyfill module.
With this information we might then try to speed up this part of the code
by altering the algorithm or making the limbo code more efficient (for
instance by moving constant calculations or addressing out of loops).
.LP
The number of sampling points is also shown. The sampling rate can be
increased with the -s option to give better granularity.
This will cause a decrease in apparent performance but increases the
accuracy of the results. The above example showed the results for all
modules sampled. We might have restricted attention to the two main
polyhedra modules instead by executing
.sp
.RS
	prof -m Polyhedra -m Polyfill wm/polyhedra
.RE
.sp
See the manual page for other options to 
.I prof
and further examples.
.NH1
Cprof
.LP
Coverage of instructions is achieved by running a special dis instruction execute routine in place of the usual one (just as the debugger does).
This routine notes down each
instruction as it is executed. The profile device then passes this information
to 
.I cprof
via the io system.
.LP
The coverage profiler is used in a similar way to the time profiler.
.sp
.RS
	cprof -m Zeros zeros 1024 2880
.RE
.sp
gives
.br
.DS
Module: Zeros(zeros.dis)	56% coverage

1	 	implement Zeros;
2	 	
3	 	include "sys.m";
4	 		sys: Sys;
5	 	include "arg.m";
6	 		arg: Arg;
7	 	include "string.m";
8	 		str: String;
9	 	include "keyring.m";
10	 	include "security.m";
11	 		random: Random;
12	 	
13	 	include "draw.m";
14	 	
15	 	Zeros: module
16	 	{
17	 		init: fn(nil: ref Draw->Context, argv: list of string);
18	 	};
19	 	
20	 	init(nil: ref Draw->Context, argv: list of string)
21	 	{
22	 		z: array of byte;
23	 		i: int;
24	+		sys = load Sys Sys->PATH;
25	+		arg = load Arg Arg->PATH;
26	+		str = load String String->PATH;
27	 	
28	+		if(sys == nil || arg == nil)
29	-			return;
30	 	
31	+		bs := 0;
32	+		n := 0;
33	+		val := 0;
34	+		rflag := 0;
35	+		arg->init(argv);
36	+		while ((c := arg->opt()) != 0)
37	-			case c {
38	-			'r' => rflag = 1;
39	-			'v' => (val, nil) = str->toint(arg->arg(), 16);
40	-			* => sys->raise(sys->sprint("fail: unknown option (%c)\n", c));
41	 			}
.DE
.DS
42	+		argv = arg->argv();
43	+		if(len argv >= 1)
44	+			bs = int hd argv;
45	 		else
46	-			bs = 1;
47	+		if (len argv >= 2)
48	+			n = int hd tl argv;
49	 		else
50	-			n = 1;
51	+		if(bs == 0 || n == 0) {
52	-			sys->fprint(sys->fildes(2), "usage: zeros [-r] [-v value] blocksize [number]\n");
53	-			sys->raise("fail: usage");
54	 		}
55	+		if (rflag) {
56	-			random = load Random Random->PATH;
57	-			if (random == nil)
58	-				sys->raise("fail: no security module\n");
59	-			z = random->randombuf(random->NotQuiteRandom, bs);
60	 		}
61	 		else {
62	+			z = array[bs] of byte;
63	+			for(i=0;i<bs;i++)
64	+				z[i] = byte val;
65	 		}
66	+		for(i=0;i<n;i++)
67	+			sys->write(sys->fildes(1), z, bs);
68	+	}

**** module dis instructions 39725 ****
.DE
.br
Here the -m option has restricted attention to the Zeros module itself.
The output shows the source line number, an indicator of coverage and
the source. The indicator is + if the line has been executed, - if
it hasn't and ? if only part of it has (for instance a loop statement that has
never had it's iteration part executed). Lines with no indicator have no
corresponding dis instructions associated with them. Another option (-f)
shows coverage frequencies instead.
.LP
An alternative to 
.I cprof
is 
.I wm/cprof
which shows the statistics graphically.
It's options are pretty much the same as those to 
.I cprof .
Unexecuted and 
partially executed lines of code are shown in colour. See the man page 
for exact details of the colouring convention
.LP
Results may be accumulated with the -r option so that multiple runs of
code can be made. The resulting statistics go into a file <xxx>.prf when
<xxx>.dis is the corresponding dis file. See the manual page for further
details on how to use this option and then review the accumulated 
results.
.NH 1
Mprof
.LP
When memory profiling, the kernel profile device associates each heap allocation with a line of limbo source and each heap deallocation with the line of
limbo source that allocated it. In this way, current memory usage and
high-water usage can be determined on a line by line basis.
.LP
Here it seems that memory usage at a particular point in the execution of
a program is more appropriate than the post-mortem approach of 
.I prof
and 
.I cprof
, so an interactive example is described (though 
.I mprof
can be
used non-interactively and 
.I prof
interactively if so desired). See the manual
pages for complete details and further examples.
.LP
To do this we execute
.sp
.RS
	mprof -b -m Polyhedra
.RE
.sp
which kicks off profiling and restricts attention to the Polyhedra module
whenever it runs. The -b simply says begin profiling. Note that no command
to execute is given to 
.I mprof
at this stage. Then run the command
.sp
.RS
	wm/polyhedra &
.RE
.sp
and interact with it. Now show memory statistics
.sp
.RS
	mprof
.RE
.sp
This gives
.br
.DS
Module: Polyhedra(/dis/wm/polyhedra.dis)

44	100	100		sys = load Sys Sys->PATH;
45	132	132		draw = load Draw Draw->PATH;
46	68	68		tk = load Tk Tk->PATH;
47	1788	1788		wmlib = load Wmlib Wmlib->PATH;
48	232	232		bufio = load Bufio Bufio->PATH;
49	68	68		math = load Math Math->PATH;
50	204	204		rand = load Rand Rand->PATH;
51	0	3504		daytime = load Daytime Daytime->PATH;
52	544	544		polyfill = load Polyfill Polyfill->PATH;
53	1824	1824		smenu = load Smenu Smenu->PATH;
86	36	36		cmdch := chan of string;
95	36	36		sync := chan of int;
96	36	36		chanθ := chan of real;
103	68	68		shade = array[NSHADES] of ref Image;
116	36	36		yieldc := chan of int;
120	36	36		sm := array[2] of ref Scrollmenu;
378	68	176			s += " (" + string p.indx + ")";
403	36	36		vec := array[2] of array of Vector;
404	740	740		vec[0] = array[V] of Vector;
405	740	740		vec[1] = array[V] of Vector;
407	36	36			norm = array[2] of array of Vector;
408	612	612			norm[0] = array[F] of Vector;
409	612	612			norm[1] = array[F] of Vector;
492	68	68		ap := array[n+1] of Point;
609	164	164		geo := ref Geom;
610	36	36		TM := array[4] of array of real;
612	272	272			TM[i] = array[4] of real;
663	8000	8000			p := ref Polyhedron;
707	740	740		p.v = array[p.V] of Vector;
710	612	612		p.f = array[p.F] of Vector;
713	132	132		p.fv = array[p.F] of array of int;
716	164	164		p.vf = array[p.V] of array of int;
729	9504	9640			return s[0: len s - 1];
750	3672	3672		a := array[n+2] of int;
768	0	136				return (n, s[i+1:]);
779	0	104				return (r, s[i+1:]);
802	0	68			s = s[1:];
806	0	72			s = s[0: ln-1];
.DE
.DS
866	0	200		cmd(mainwin, ".f1.txt configure -text {" + s + "}");
874	0	356		labs := array[n] of string;
881	0	5128			labs[i++] = string q.indx + " " + name;
884	0	68		cmd(top, mname + " configure -borderwidth 3");
920	0	104			cmd(win, ". configure -height " + string (scrsize.y - bd * 2));
934	0	244		cmd(win, ". configure -x " + string actr.min.x + " -y " + string actr.min.y);

Module totals	31416	33984
.DE
.br
We get the source line number, the amount of memory in bytes
currently allocated on that line, the high-water mark in bytes and then the source. Thus
loading the Sys module on line 44 used 100 bytes and this memory is
still allocated. Loading Daytime on line 51 used 3504 bytes but this is now
released (because the module pointer is set to nil in the source and the
memory has been reclaimed). The string concatenation on line 378 currently uses 68 bytes 
but at it's worst it was 176 bytes.
.LP
Further interaction with wm/polyhedra can now be done and memory
statistics reviewed before the
profiling session is ended, throwing away the accumulated memory
statistics inside the kernel with
.sp
.RS
	mprof -c
.RE
.sp
The -c option simply says cease profiling.
.LP
An alternative to 
.I mprof
is 
.I wm/mprof
which shows the statistics graphically.
It's options are pretty much the same as those to 
.I mprof .
Lines of code
which have allocated more of the memory are shown in darker shades of red.
.NH 1
Manual pages
.LP
Further information and other examples are given in the following
manual pages :-
.sp
.RS
.I cprof(1)
.br
.I mprof(1)
.br
.I prof(1)
.br
.I wm-cprof(1)
.br
.I wm-mprof(1)
.RE
.sp
For the lower level library module interface to these profilers
.sp
.RS
.I prof(2)
.RE
.sp
For the kernel profile device which gathers timing, coverage and memory
statistics
.sp
.RS
.I prof(3)
.RE
.br
.NH 1
Sources
.LP
The relevant sources are
.sp
.RS
.CW /module/profile.m
.br
.CW /appl/lib/profile.b
.br
.CW /appl/cmd/cprof.b
.br
.CW /appl/cmd/mprof.b
.br
.CW /appl/cmd/prof.b
.br
.CW /appl/wm/cprof.b
.br
.CW /appl/wm/mprof.b
.br
.CW /emu/devprof.c
.br
.CW /os/port/devprof.c
.RE
.NH 1
Addendum
.LP
A gui version of
.I prof
has been added for completeness. See the manual page
.I wm-prof(1)
and the source
.CW /appl/wm/prof.b .