[M3devel] performance problems in CM3 Text operations on AMD64_LINUX
Jay K
jay.krell at cornell.edu
Fri Jan 10 09:22:25 CET 2014
Ok either, way, but here is another point.
In some of our code we do:
CONST DEBUG = FALSE;
IF FALSE THEN xxEND;
It will tend to be optimized.Furthermore, syscalls -- even if not optimized,IF FALSE is going to be way faster than a syscall.
Also, gettimeofday shouldn't necessarily be a syscall.
But/and there are/should-be faster ways of getting time for benchmarking.
On Windows, GetTickCount and GetTickCount64 just read a value out of memory.Even GetSystemTimeAsFileTime is just a read of a 64bit value from memory.(The memory is read only, at the same address in all processes, and updated bythe kernel.)x86 has rdtsc, a 64bit hardware cycle counter readable in one instruction.
We should consider researching analogs to rdtsc across mips/powerpc/sparc/arm andprovide something in m3core. I think.Add it to the backend interface likely, for inlining.or maybe it is worth a function call, then just use #ifdefed C.
Maybe Posix provides something that tends to be like this?
Mika, how did you find this?Control-c every so often and look at callstack?Or a lot of stepping?
- Jay
: Thu, 9 Jan 2014 15:01:55 -0600
> From: rodney_bates at lcwb.coop
> To: mika at async.caltech.edu
> CC: m3devel at elegosoft.com
> Subject: Re: [M3devel] performance problems in CM3 Text operations on AMD64_LINUX
>
>
>
> On 01/04/2014 12:55 PM, mika at async.caltech.edu wrote:
> > I have this program that generates Modula-3 code from XML.
> >
> > It uses Wr.PutText a lot, on small strings.
> >
> > It runs 10 times faster under PM3 on FreeBSD/i386 on a 2 GHz old Pentium
> > III type computer than under CM3 (at head) on Debian/amd64 on a 4 GHz
> > Core i7 4960X.
> >
> > (176)rover:~/t/calarm/fix/gen/src>time ../FreeBSD4/genfix -o ../../fix42/src -T ../templates -L fix42 FIX42.xml
> > 0.419u 0.089s 0:00.70 70.0% 471+27608k 0+546io 0pf+0w
> >
> > (29)truffles:~/t/calarm/fix/gen/src>time ../AMD64_LINUX/genfix -o ../../fix42/src -T ../templates -L fix42 FIX42.xml
> > 1.952u 0.232s 0:05.40 40.3% 0+0k 0+11504io 0pf+0w
> >
> > I believe the issue is that Text.m3 is making a ton of system calls!
> >
> > Here's a typical traceback:
> >
> > (gdb) where
> > #0 0x00002aaaaaacb770 in ?? ()
> > #1 0x00002aaaaaacba1b in gettimeofday ()
> > #2 0x00002aaaac468f8a in gettimeofday () at ../sysdeps/unix/sysv/linux/x86_64/gettimeofday.S:37
> > #3 0x000000000048c10c in TimePosix__Now () at ../src/time/POSIX/TimePosixC.c:50
> > #4 0x000000000048b4d2 in Time__Now () at ../src/time/POSIX/TimePosix.m3:14
> > #5 0x000000000049cbcf in TextStats__NoteFinished (M3_ESffkp_o=<error reading variable>) at ../src/text/TextStats.m3:64
> > #6 0x00000000004921de in Text__SetChars (M3_CKMnXU_a=<error reading variable>, M3_Bd56fi_t=<error reading variable>,
> > M3_Cwb5VA_start=<error reading variable>) at ../src/text/Text.m3:512
> > #7 0x0000000000441a26 in UnsafeWr__FastPutText (M3_BxxOH1_wr=<error reading variable>, M3_Bd56fi_t=<error reading variable>) at ../src/rw/Wr.m3:100
> > #8 0x000000000044192f in Wr__PutText (M3_BxxOH1_wr=<error reading variable>, M3_Bd56fi_t=<error reading variable>) at ../src/rw/Wr.m3:89
> > #9 0x000000000041627f in M3Syntax__TextSetPutList (M3_CT76zs_s=<error reading variable>, M3_BxxOH1_wr=<error reading variable>,
> > M3_AicXUJ_quoted=<error reading variable>) at ../src/M3Syntax.m3:129
> > #10 0x00000000004169cf in M3Syntax__EnumToText (M3_Bqnwti_e=<error reading variable>, M3_ACGgGY_env=<error reading variable>) at ../src/M3Syntax.m3:173
> > #11 0x000000000041791a in M3Syntax__UnstrToText (M3_AIPC2K_u=<error reading variable>, M3_ACGgGY_env=<error reading variable>) at ../src/M3Syntax.m3:295
> > #12 0x0000000000414ad4 in M3File__DumpInterface (M3_CjySNq_t=<error reading variable>, M3_Bd56fi_dirPath=<error reading variable>,
> > M3_AicXUJ_updateMakefile=<error reading variable>) at ../src/M3File.m3:78
> > #13 0x000000000040f3f9 in Fields__Do__MakeFieldsFiles.1150 () at ../AMD64_LINUX/Fields.m3:526
> > #14 0x000000000040a713 in Fields__Do (M3_DsEykq_xml=<error reading variable>, M3_Bd56fi_versionString=<error reading variable>,
> > M3_Bd56fi_beginStringV=<error reading variable>, M3_D9m5ya_fieldTbl=<error reading variable>, M3_Bd56fi_Dir=<error reading variable>,
> > M3_Bd56fi_SrcDir=<error reading variable>, M3_AZx9O5_binaryFields=<error reading variable>, M3_CebKnt_specialFields=<error reading variable>,
> > M3_CT76zs_checkedTypes=<error reading variable>) at ../AMD64_LINUX/Fields.m3:538
> > #15 0x0000000000413948 in Main_M3 (M3_AcxOUs_mode=<error reading variable>) at ../AMD64_LINUX/Main.m3:431
> > #16 0x00000000004742ad in RTLinker__RunMainBody (M3_DjPxE3_m=<error reading variable>) at ../src/runtime/common/RTLinker.m3:408
> > #17 0x0000000000473638 in RTLinker__AddUnitI (M3_DjPxE3_m=<error reading variable>) at ../src/runtime/common/RTLinker.m3:115
> > #18 0x00000000004736cc in RTLinker__AddUnit (M3_DjPxE5_b=<error reading variable>) at ../src/runtime/common/RTLinker.m3:124
> > #19 0x0000000000405df8 in main (argc=8, argv=0x7fffffffe2b8, envp=0x7fffffffe300) at _m3main.c:22
> > (gdb)
> >
> > I believe all the TextStats operations are killing the performance. Can we remove them?
> >
> > Here is SetChars:
> >
> > PROCEDURE SetChars (VAR a: ARRAY OF CHAR; t: T; start: CARDINAL) =
> > BEGIN
> > TextStats.NoteGround (TextStats.Op.SetChars);
> > TextStats.NoteGround (TextStats.Op.get_chars);
> > t.get_chars (a, start);
> > TextStats.NoteFinished (TextStats.Op.get_chars);
> > TextStats.NoteFinished (TextStats.Op.SetChars);
> > END SetChars;
> >
> > I'm not sure how best to do this. Removing them without killing functionality might not be
> > easy. Looks like a case for conditional compilation.
>
> Now the instrumentation calls are commented out with (*47 ... 74*), which could be reinstated
> by string search & replace "(*47" -> "(*47*)", etc., and removed again similarly. These are
> only needed for performance study of the algorithms, which likely won't be done often.
>
> >
> > Maybe generate the .m3s from some other source, with and without the TextStats, depending on
> > a setting in m3makefile?
> >
> > Second best would be to have a boolean flag in Text.m3 so the TextStats never get called.
> >
> > Mika
> >
> >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://m3lists.elegosoft.com/pipermail/m3devel/attachments/20140110/fe64b4e8/attachment-0002.html>
More information about the M3devel
mailing list