parallelism in Sage: just use value of 'MAKE'

[jhpalmieri]

The various parallel aspects of Sage should be controlled by setting the -j (possible also -l) flags in MAKE or MAKEFLAGS. That is, if MAKE='make -j16', then

• running make will build spkg's in parallel, using 16 processes (this was done in #11959). This is standard make behaviour, but we need to patch spkg/standard/deps to ensure that make recognizes that we are doing a recursive make.

• running make ptestlong or sage -tp 0 <files> will doctest in parallel using 16 threads. If the -j flag in MAKE is not set, then determine the number of threads as before: min(8, cpu_count()).

• running ./sage -b will build the Sage library using 16 threads. If the -j flag in MAKE is not set, then use only 1 thread.

Testing this ticket: you can set the environment variable SAGE_NUM_CORES to the number of cores you want to pretend to have. For example, running

SAGE_NUM_CORES=24 make ptestlong

should run 8 threads (see sage-num-threads.py; this is undocumented because the only purpose I see is for testing this ticket).

Notes: With the patches applied, building spkgs in parallel works well, except for race conditions in:

• python (#12096)
• singular (#12137)
• zlib (#12138)
• mpir (#12139)

and a "jobserver unavailable" warning in:

• ntl
• singular (#12137)
• rubiks

Apply:

1. 12016-root.patch​ to the SAGE_ROOT repository.
2. 12016-base.patch​ to spkg/base.
3. 12016-scripts.patch​ and trac_12016-scripts-ref.patch​ to the SCRIPTS repository.
4. 12016-sage.patch​ to the Sage library.

See also: #6495 to implement the same behavior for doc building.

[jdemeyer]

We should remove NUM_THREADS from the top-level Makefile.

[jdemeyer]

John, with your solution there is a lot of code duplication (determining the number of threads is done in 3 places, potentially in 3 different ways). How about having code in sage-sage or sage-env to determine the number of threads and saving it in an environment variable SAGE_NUM_PROCESSES (which the user could set by hand; if not set, the value comes from MAKE or MAKEFLAGS; if no -j option is given, set to 1).

[jhpalmieri]

Replying to jdemeyer:

John, with your solution there is a lot of code duplication (determining the number of threads is done in 3 places, potentially in 3 different ways). How about having code in sage-sage or sage-env to determine the number of threads and saving it in an environment variable SAGE_NUM_PROCESSES

Sounds okay.

(which the user could set by hand; if not set, the value comes from MAKE or MAKEFLAGS; if no -j option is given, set to 1).

If you run "sage -tp <files>", should you use 1 process or more than 1? The "-tp" option means "parallel", so perhaps the default should be more than 1 in this case. In other cases (like docbuilding, for example), the default should be 1.

[jhpalmieri]

For something like make -j16 ptestlong, how do we recover the number 16? If I execute this command (with MAKE unset), I see

MAKEFLAGS= --jobserver-fds=3,4 -j
MFLAGS=- --jobserver-fds=3,4 -j

but I don't see "16" anywhere in the listing of the environment variables.

[jdemeyer]

Replying to jhpalmieri:

Replying to jdemeyer:

John, with your solution there is a lot of code duplication (determining the number of threads is done in 3 places, potentially in 3 different ways). How about having code in sage-sage or sage-env to determine the number of threads and saving it in an environment variable SAGE_NUM_PROCESSES

Sounds okay.

(which the user could set by hand; if not set, the value comes from MAKE or MAKEFLAGS; if no -j option is given, set to 1).

If you run "sage -tp <files>", should you use 1 process or more than 1? The "-tp" option means "parallel", so perhaps the default should be more than 1 in this case. In other cases (like docbuilding, for example), the default should be 1.

Sure, that is what I meant. We should compute the value once, but in sage -tp we can still decide to use the number of processes.

[jdemeyer]

Replying to jhpalmieri:

For something like make -j16 ptestlong, how do we recover the number 16? If I execute this command (with MAKE unset), I see

MAKEFLAGS= --jobserver-fds=3,4 -j
MFLAGS=- --jobserver-fds=3,4 -j

but I don't see "16" anywhere in the listing of the environment variables.

You are right. I had not tried this before. So let's scrap that idea.

[jhpalmieri]

Here are new patches. These use SAGE_NUM_THREADS if it is set, and otherwise try to extract a number from MAKE. (My method for doing this is probably not ideal, but the options This is done in sage-env. Running sage -b should use this setting now, also.

I don't know how to get the number of threads from

make -j16 ptestlong

so I removed that from the "to do" list in the ticket description.

In the file sage-ptest, I removed the "FIXME" comment in

    try:
        # FIXME: Nice, but <NUMTHREADS> should immediately follow '-tp' etc.,
        #        i.e., be the next argument. We might have file or directory
        #        names that properly convert to an int...
        numthreads = int(argv[1])
        infiles = argv[2:]
    except ValueError: # can't convert first arg to an integer: arg was probably omitted
        numthreads = 1

The script sage-ptest doesn't get a "tp" argument; it is instead called by sage-sage, and the way it is called, the first argument to sage-ptest is precisely what ever came after "-tp". So I don't think anything needs fixing. If we ever rewrite sage-sage (#21) to properly parse arguments, we can make sure that "-tp" has a default numerical argument of zero.

[jdemeyer]

1) If you are going to use the string "auto" for automatic, you might as well use "infinite" for infinite, instead of zero.

1b) Alternatively: use 0 for automatic (as is sage -tp 0) and 999999 for unlimited. This would mean less special-case code, since a value like 999999 is more than what a user would normally specify (for the forseeable future).

2) In sage-ptest, unlimited really should be unlimited. Not max(8, # of cpus).

3) We should also do the following long-needed fix here: setting MAKE to make -j16 is very standard in Sage circles, but not actually the prefered way according to the GNU make folks. One really should use MAKEFLAGS instead (similar to the distinction between CC and CFLAGS). This is why you often see an error like "make -jN forced in sub-make. Disabling job server mode" (freely quoted from my mind). So, when MAKEFLAGS exists, assume that make understands the flags and do not pass flags in MAKE.

4) Why did you change

sage-build "$@" || exit $?

to

sage-build "$@"

in the sage_build() function in sage-sage?

5) You reverted a lot of changes that I made to doc/en/developer/doctesting.rst. Why? I actually tried all the examples in the documentation and pasted the exact output I got (on sage.math.washington.edu). Surely, this is better than keeping the outdated (and in many cases totally wrong) output.

I am planning to work further on this, so don't change any code yet. But please give your opinion.

[jhpalmieri]

Replying to jdemeyer:

1) If you are going to use the string "auto" for automatic, you might as well use "infinite" for infinite, instead of zero.

1b) Alternatively: use 0 for automatic (as is sage -tp 0) and 999999 for unlimited. This would mean less special-case code, since a value like 999999 is more than what a user would normally specify (for the forseeable future).

Sounds good to me.

2) In sage-ptest, unlimited really should be unlimited. Not max(8, # of cpus).

Okay.

3) We should also do the following long-needed fix here: setting MAKE to make -j16 is very standard in Sage circles, but not actually the prefered way according to the GNU make folks. One really should use MAKEFLAGS instead (similar to the distinction between CC and CFLAGS). This is why you often see an error like "make -jN forced in sub-make. Disabling job server mode" (freely quoted from my mind). So, when MAKEFLAGS exists, assume that make understands the flags and do not pass flags in MAKE.

I'm willing to try that, especially if you write the patch instead of me :)

4) Why did you change

sage-build "$@" || exit $?

to

sage-build "$@"

in the sage_build() function in sage-sage?

That was a mistake.

5) You reverted a lot of changes that I made to doc/en/developer/doctesting.rst. Why? I actually tried all the examples in the documentation and pasted the exact output I got (on sage.math.washington.edu). Surely, this is better than keeping the outdated (and in many cases totally wrong) output.

Some of them I disagreed with, like the complete removal of the section "Beyond the Sage library". So I started from scratch, at which point I just put in the changes that I felt were relevant to the ticket or easy for me to change. Probably I should have started with your patch and added the section (with modifications) back in.

It looks like #9739 broke doctesting of .sage files. We should fix that (not on this ticket).

[jhpalmieri]

Replying to jhpalmieri:

It looks like #9739 broke doctesting of .sage files. We should fix that (not on this ticket).

See #12069.

[jdemeyer]

Replying to jhpalmieri:

Some of them I disagreed with, like the complete removal of the section "Beyond the Sage library".

I removed that because it totally didn't work. But this is probably #12069. How about we leave the last section of the documentation alone in this ticket but then change the documentation in #12069?

[jhpalmieri]

Replying to jdemeyer:

Replying to jhpalmieri:

Some of them I disagreed with, like the complete removal of the section "Beyond the Sage library".

I removed that because it totally didn't work. But this is probably #12069. How about we leave the last section of the documentation alone in this ticket but then change the documentation in #12069?

Okay, sounds fine to me.

[jdemeyer]

The essence of the patch should be there. I'm not claiming it works, I still need to test many things.

[jhpalmieri]

In spkg/standard/deps, why prefix most of the rules with "+"?

[jdemeyer]

Replying to jhpalmieri:

In spkg/standard/deps, why prefix most of the rules with "+"?

To mark the rule as "recursive", see ​http://www.gnu.org/software/make/manual/make.html#Recursion. Otherwise you get lots of warnings like

make[2]: warning: jobserver unavailable: using -j1.  Add `+' to parent make rule.

[jdemeyer]

With these patches applied, I often (not always) get the following doctest error in sage0.py:

sage -t  -force_lib devel/sage/sage/interfaces/sage0.py
**********************************************************************
File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/devel/sage-main/sage/interfaces/sage0.py", line 448:
    sage: F == sage0(F)._sage_()
Exception raised:
    Traceback (most recent call last):
      File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/local/bin/ncadoctest.py", line 1231, in run_one_test
        self.run_one_example(test, example, filename, compileflags)
      File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/local/bin/sagedoctest.py", line 38, in run_one_example
        OrigDocTestRunner.run_one_example(self, test, example, filename, compileflags)
      File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/local/bin/ncadoctest.py", line 1172, in run_one_example
        compileflags, 1) in test.globs
      File "<doctest __main__.example_20[4]>", line 1, in <module>
        F == sage0(F)._sage_()###line 448:
    sage: F == sage0(F)._sage_()
      File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/local/lib/python/site-packages/sage/interfaces/sage0.py", line 458, in _sage_
        return load(P._local_tmpfile())
      File "sage_object.pyx", line 775, in sage.structure.sage_object.load (sage/structure/sage_object.c:7937)
    IOError: [Errno 2] No such file or directory: '/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/home/.sage//temp/sage.math.washington.edu/4956//interface//tmp5116.sobj'
**********************************************************************
[...]
**********************************************************************
File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/devel/sage-main/sage/interfaces/sage0.py", line 547:
    sage: sage0_version() == version()
Expected:
    True
Got:
    False
**********************************************************************
[...]
**********************************************************************
File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/devel/sage-main/sage/interfaces/sage0.py", line 176:
    sage: sage0('factor(2^157-1)')
Expected:
    852133201 * 60726444167 * 1654058017289 * 2134387368610417
Got:
    <BLANKLINE>
**********************************************************************
File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/devel/sage-main/sage/interfaces/sage0.py", line 178:
    sage: print "ignore this";  sage0.cputime()     # random output
Exception raised:
    Traceback (most recent call last):
      File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/local/bin/ncadoctest.py", line 1231, in run_one_test
        self.run_one_example(test, example, filename, compileflags)
      File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/local/bin/sagedoctest.py", line 38, in run_one_example
        OrigDocTestRunner.run_one_example(self, test, example, filename, compileflags)
      File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/local/bin/ncadoctest.py", line 1172, in run_one_example
        compileflags, 1) in test.globs
      File "<doctest __main__.example_3[4]>", line 1, in <module>
        print "ignore this";  sage0.cputime()     # random output###line 178:
    sage: print "ignore this";  sage0.cputime()     # random output
      File "/mnt/usb1/scratch/jdemeyer/merger/sage-4.8.alpha3-make-jobs/local/lib/python/site-packages/sage/interfaces/sage0.py", line 185, in cputime
        return float(s)
    ValueError: invalid literal for float(): 852133201 * 60726444167 * 1654058017289 * 2134387368610417
**********************************************************************

[jhpalmieri]

I installed these patches on sage.math and set MAKE='make -j12'. Then I ran make ptestlong, and the log file for the testing says "Doctesting 2857 files doing 2 jobs in parallel". Running ./sage -tp --long ... uses 12 threads, as it should.

If I remove the redirection of stderr to null in sage-env, when calling sage-num-threads, I see this:

Traceback (most recent call last):
  File "/mnt/usb1/scratch/palmieri/12016/test1/sage-4.8.X/local/bin/sage-num-threads.py", line 144, in <module>
    print "%i %i %i"%num_threads()
  File "/mnt/usb1/scratch/palmieri/12016/test1/sage-4.8.X/local/bin/sage-num-threads.py", line 95, in num_threads
    if MAKEFLAGS[0] != '-':
IndexError: string index out of range

This seems to fix the problem for me:

sage-num-threads.py

@@ -92 +92 @@
     try:
         # Prepend hyphen to MAKEFLAGS if it does not start with one
         MAKEFLAGS=os.environ["MAKEFLAGS"]
-        if MAKEFLAGS[0] != '-':
+        if len(MAKEFLAGS) > 0 and MAKEFLAGS[0] != '-':
             MAKEFLAGS = '-' + MAKEFLAGS
         # In MAKEFLAGS, "-j" does not mean unlimited.  It probably
         # means an inherited number of jobs, let us use 2 for safety.

[jhpalmieri]

Another patch:

setup.py

@@ -242 +242 @@
     WARNING: commands are run roughly in order, but of course successive
     commands may be run at the same time.
     """
-    print "Execute %s commands (using %s threads)"%(len(command_list), min(len(command_list),nthreads))
+    nthreads = min(len(command_list),nthreads)
+    print "Execute %s commands (using %s threads)"%(len(command_list), nthreads)
     from multiprocessing import Pool
     import twisted.persisted.styles #doing this import will allow instancemethods to be pickable
     p = Pool(nthreads)

Without this one, if you do export MAKE='make -j' and then make ptestlong, although it says it's using 0 threads, it actually starts up 999999 threads (to do nothing, as far as I can tell).

I get lots of doctest failures and warnings using make -j, along the lines of OSError: [Errno 11] Resource temporarily unavailable and /bin/sh: Cannot fork, but I guess this is to be expected? Maybe we should put some sort of cap on the number of threads to use for doctesting?

[jhpalmieri]

Replying to jdemeyer:

With these patches applied, I often (not always) get the following doctest error in sage0.py:

How many threads were you using? I haven't seen this recently, but I think I've seen this sort of error before unrelated to this ticket. I'm not sure I would consider it an obstacle for this ticket. It would be nice to track down the problem, though.

[jdemeyer]

Thanks for the feedback, these should be easy fixes.

[jhpalmieri]

For this ticket or a follow-up: it would be nice to be able to set MAKE=make -j -l30, for example, and have tests pass. Building this way works, but I get lots of doctest failures as ​mentioned above. Maybe we could read the argument for -l for a cap on the threads for doctesting? Or use the number of cpus for this cap?

[jdemeyer]

Replying to jhpalmieri:

Maybe we could read the argument for -l for a cap on the threads for doctesting?

Should be possible, in sage-num-threads.py.

[jdemeyer]

Since execute_list_of_commands_in_serial is essentially a special case of execute_list_of_commands_in_parallel, I see no reason to keep the former.

[jdemeyer]

New patch up, implements support for "-l" option to make, fixes and simplifies setup.py.

[jhpalmieri]

This looks good to me. Is it ready for review? Am I allowed to review it since I wrote early drafts of some of the patches?

For a future ticket, it would be nice if you could set MAKE='make -j -lN', for some reasonable choice of N, and have it work. When I try this, I have problems with the following spkgs, and I'm not sure why:

• zlib on OS X (2 cores) fails most of the time with MAKE='make -j -l3'. ​Here's a log.
• singular on sage.math fails all of the time, I think, with MAKE='make -j -l30'. ​Here's a log.

(The Sage spkg used to fail before the latest round of patches using the -l setting for a cap on the number of threads, and the same goes for parallel doctesting. Setting MAKE='make -j' still causes these failures. Should we just regard this setting as too dangerous, or try to stop the failures by putting a cap on the number of processes if -j is present but set to "unlimited" and -l is missing?)

These failures are not related to this ticket; they fail with or without the patches. But the ticket makes it more appealing to just set MAKE as above.

[jdemeyer]

Replying to jhpalmieri:

This looks good to me. Is it ready for review?

Yes, it is. I just didn't want to put "needs review" because I have not really tested it.

Am I allowed to review it since I wrote early drafts of some of the patches?

I would say yes, since I certainly looked at all your code. So consider all your code to be positively reviewed by me.

For a future ticket, it would be nice if you could set MAKE='make -j -lN', for some reasonable choice of N, and have it work. When I try this, I have problems with the following spkgs, and I'm not sure why:

It should work.

• zlib on OS X (2 cores) fails most of the time with MAKE='make -j -l3'. ​Here's a log.
• singular on sage.math fails all of the time, I think, with MAKE='make -j -l30'. ​Here's a log.

Question for both cases: does MAKE="make -jN" work for various values of N? Because I don't see a fundamental difference between "make -jN" and "make -j -lN".

(The Sage spkg used to fail before the latest round of patches using the -l setting for a cap on the number of threads, and the same goes for parallel doctesting. Setting MAKE='make -j' still causes these failures. Should we just regard this setting as too dangerous?

Yes. Since "make" will simply run as many threads as it can, I think Sage should do the same, no matter how stupid that is.

By the way, I would really like to merge this in the sage-4.8 release, because it cleans up some stuff which will also help future tickets like #11073 (which hopefully will be merged in sage-5.0).

[jdemeyer]

Replying to jhpalmieri:

• zlib on OS X (2 cores) fails most of the time with MAKE='make -j -l3'. ​Here's a log.
• singular on sage.math fails all of the time, I think, with MAKE='make -j -l30'. ​Here's a log.

Since my patches properly implement parallel building, it also means that more packages are actually being built in parallel. So I think we are simply triggering bugs in the various packages. For example, I never had problems with Python before, but I did have problems with this patch (fixed in #12096).

I cannot explain why "make -j -lN" would fail but "make -jN" would work.

[jdemeyer]

Replying to jhpalmieri:

• singular on sage.math fails all of the time, I think, with MAKE='make -j -l30'. ​Here's a log.

Well, singular is in the list of fishy packages, see the ticket description.

[jhpalmieri]

Just for fun, I modified deps so singular would build all by itself in the build process (I made it depend on linbox and scipy, so it was the last package to be built before the sage package). Then it built fine using make -j -l30.

[jdemeyer]

I think it is truly a coincidence that "make -j -lN" fails. I managed to make singular fail with just "make -jN", hopefully fixed by #12138.

[jdemeyer]

Replying to jhpalmieri:

This looks good to me. Is it ready for review? Am I allowed to review it since I wrote early drafts of some of the patches?

For a future ticket, it would be nice if you could set MAKE='make -j -lN', for some reasonable choice of N, and have it work. When I try this, I have problems with the following spkgs, and I'm not sure why:

• zlib on OS X (2 cores) fails most of the time with MAKE='make -j -l3'. ​Here's a log.
• singular on sage.math fails all of the time, I think, with MAKE='make -j -l30'. ​Here's a log.

Are these about builds of the total Sage source in which these fail, or are these separate installs like sage -f ...?

[jdemeyer]

When testing with sage -f, the proper way to test is using

MAKEFLAGS="j50" ./sage -f ...

[jhpalmieri]

In the following lines from sage-spkg

# Handle -n, -t, -q options for recursive make 
# See Trac #12016. 
if echo "$MAKE $MAKEFLAGS -$MAKEFLAGS" |grep -e ' -[A-Za-z]*[qnt]' >/dev/null; then 
    if echo "$MAKE $MAKEFLAGS -$MAKEFLAGS" |grep -e ' -[A-Za-z]*q' >/dev/null; then 
        exit 1 
    else 
        exit 0 
    fi 
fi 

do we also need to handle the long versions? (I don't think so, but I thought I would ask.)

More importantly, on OpenSolaris, or at least on David Kirkby's machine hawk, the default 'grep' command doesn't take a -e option. Can we just omit it? The command still seems to function on sage.math, on OS X, and on OpenSolaris.

[jhpalmieri]

I cannot explain why "make -j -lN" would fail but "make -jN" would work.

One reason is that make -j -lN puts a limit on starting new processes, and that might be what's causing the problems. I could force the old zlib spkg to fail on sage.math by running MAKEFLAGS='j -l2' ./sage -f ... but not with MAKEFLAGS='j -l30' .... I don't know if setting MAKE="$MAKE -j1 -l in spkg-install is the right way to fix this for problematic spkgs (like singular?), but it might be worth trying.

[jdemeyer]

Replying to jhpalmieri:

In the following lines from sage-spkg

# Handle -n, -t, -q options for recursive make 
# See Trac #12016. 
if echo "$MAKE $MAKEFLAGS -$MAKEFLAGS" |grep -e ' -[A-Za-z]*[qnt]' >/dev/null; then 
    if echo "$MAKE $MAKEFLAGS -$MAKEFLAGS" |grep -e ' -[A-Za-z]*q' >/dev/null; then 
        exit 1 
    else 
        exit 0 
    fi 
fi 

do we also need to handle the long versions? (I don't think so, but I thought I would ask.)

Well, this would only be needed if the user does something very silly like

MAKE="make --dry-run" ./sage -f ...

More importantly, on OpenSolaris, or at least on David Kirkby's machine hawk, the default 'grep' command doesn't take a -e option. Can we just omit it?

Probably yes, but it might be safer to replace the leading space by a [ ].

[jhpalmieri]

I'm happy with this except for a few small changes I want to make in sage-num-threads.py: we should use subprocess instead of popen, since popen has been deprecated. Also, we should catch errors if sysctl fails to run -- it's not present on all platforms. Finally, we might as well search for max-load in addition to load-average. See the referee patch. If you're happy with that, the whole thing can get a positive review.

[jhpalmieri]

scripts repo