ikol's profile - activity

The parallel Sage interface PSage()http://doc.sagemath.org/html/en/refer... works fine with the given example, but I have trouble with a more complex case, which would be a typical application of this very useful feature. The following code works exactly as advertised:

>>> v = [ PSage() for _ in range(5)]
>>> w = [x('factor(2**%s-1)'% randint(250,310)) for x in v]
>>> print w
[127 * 13367 * 164511353 * 17137716527 * 51954390877748655744256192963206220919272895548843817842228913,
 ,
 <<currently executing code>>,
 3 * 5^2 * 11 * 31 * 41 * 53 * 131 * 157 * 521 * 1613 * 2731 * 8191 * 51481 * 409891 * 7623851 * 34110701 * 108140989558681 * 145295143558111,
 ]
[127 * 13367 * 164511353 * 17137716527 * 51954390877748655744256192963206220919272895548843817842228913,
 7 * 73 * 16183 * 34039 * 1437967 * 2147483647 * 833732508401263 * 658812288653553079 * 2034439836951867299888617,
 <<currently executing code>>,
 3 * 5^2 * 11 * 31 * 41 * 53 * 131 * 157 * 521 * 1613 * 2731 * 8191 * 51481 * 409891 * 7623851 * 34110701 * 108140989558681 * 145295143558111,
 ]
[127 * 13367 * 164511353 * 17137716527 * 51954390877748655744256192963206220919272895548843817842228913,
 7 * 73 * 16183 * 34039 * 1437967 * 2147483647 * 833732508401263 * 658812288653553079 * 2034439836951867299888617,
 <<currently executing code>>,
 3 * 5^2 * 11 * 31 * 41 * 53 * 131 * 157 * 521 * 1613 * 2731 * 8191 * 51481 * 409891 * 7623851 * 34110701 * 108140989558681 * 145295143558111,
 7 * 78903841 * 28753302853087 * 618970019642690137449562111 * 24124332437713924084267316537353]
[127 * 13367 * 164511353 * 17137716527 * 51954390877748655744256192963206220919272895548843817842228913,
 7 * 73 * 16183 * 34039 * 1437967 * 2147483647 * 833732508401263 * 658812288653553079 * 2034439836951867299888617,
 131071 * 12761663 * 179058312604392742511009 * 3320934994356628805321733520790947608989420068445023,
 3 * 5^2 * 11 * 31 * 41 * 53 * 131 * 157 * 521 * 1613 * 2731 * 8191 * 51481 * 409891 * 7623851 * 34110701 * 108140989558681 * 145295143558111,
 7 * 78903841 * 28753302853087 * 618970019642690137449562111 * 24124332437713924084267316537353]

Printing w repeatedly shows the progress of the five factorizations running in parallel (monitoring it looking at top is showing 5 sage/python jobs running simultaneously). The following example is global optimization, starting from 5 different starting points using the differential evolution algorithm available in SciPy. The setup is more complex but still, only a single command string is passed to PSage().

>>> v = [ PSage() for _ in range(5)]
>>> w = [x('from scipy.optimize import rosen, differential_evolution; differential_evolution(rosen, [(0,2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2)])') for x in v]
>>> print w
[Sage, Sage, Sage, Sage, Sage]

Apparently, something is wrong here, it doesn't work. But why? Let's see what happens with the serial Sage interpreter Sage().

>>> s = Sage()
>>> t = s('from scipy.optimize import rosen, differential_evolution;differential_evolution(rosen,[(0,2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2)])')
>>> print t
Sage

Looks like the same problem. However, Sage() can be made to work using the eval() method.

>>> s = Sage()
>>> t = s.eval('from scipy.optimize import rosen, differential_evolution;differential_evolution(rosen,[(0,2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2)])')
>>> print t
fun: 1.2785524204717224e-18

PSage() also has an eval() method, which AFAIK uses Sage().eval() internally, but unfortunately, it doesn't work.

>>> v = [ PSage() for _ in range(5)]
>>> w = [x.eval('from scipy.optimize import rosen, differential_evolution; differential_evolution(rosen, [(0,2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2), (0, 2)])') for x in v]
>>> print w
['<<currently executing code>>', '<<currently
executing code>>', '<<currently executing code>>',
'<<currently executing code>>', '<<currently executing
code>>']

Based on what I see looking at top the five optimization jobs do run in parallel, but even after they finish, no matter how many times I print w all I get is <<currently executing code>>.

The bottom line is that PSage() is either not working at all without using the eval() method, or, it seems to be working with the eval() method, but somehow is stuck in a bad internal state and never producing the output. Any comment is highly appreciated, it would be great to get this to work consistently.

It turns out that the problem has nothing to do with scalar vs. vector variable. The bottom line is that PSage() takes a single STRING as parameter enclosed in quotes. The following (somewhat awkward) modification makes the parallel interface work properly.

from scipy.optimize import rosen

x0 = [1.3, 0.7, 0.8, 1.9, 1.2]
rosen(x0)
848.22000000000003

v = [PSage() for _ in range(3)]
w = [x(eval('rosen(%s)'% str(x0))) for x in v]
w
[848.22, 848.22, 848.22]

Unfortunately, using eval() does not solve the problem. Because the example here is instantaneous to compute, I haven't realized that the calculations were, in fact, executed sequentially and not in parallel. After looking at the PSage() code it is clear that the argument must be a single string and the fact that this example "worked" although not in parallel, must be some artifact of using an explicit eval(). Mea culpa, HOWEVER, I created a better example and opened a new ticket, because PSage() would be an exteremely good tool for this kind of thing.

I am using the PSage()parallel interpreter as described here http://doc.sagemath.org/html/en/refer.... The interface works fine to evaluate multiple instances of function calls to functions of one or more scalar variables, however, if a function variable is array-like, the interface doesn't seem to work any more. I wonder if any special quoting is necessary for array-like arguments. Here is a minimal example. For reference, I copy the example from the above page, which works just fine, factor() has a single scalar variable.

v = [PSage() for _ in range(3)]
w = [x('factor(2^%s-1)'% randint(250,310)) for x in v]
w
[4057 * 8191 * 6740339310641 * 3340762283952395329506327023033,
 31 * 13367 * 2940521 * 164511353 * 70171342151 *
3655725065508797181674078959681,
 31 * 13367 * 2940521 * 164511353 * 70171342151 *
3655725065508797181674078959681]

However, the rosen() function with a single array-like/vector argument doesn't seem to work in the parallel interface. (The example below just calculates the same function value three times, but that is not the point here.)

from scipy.optimize import rosen

x0 = [1.3, 0.7, 0.8, 1.9, 1.2]
rosen(x0)
848.22000000000003

v = [PSage() for _ in range(3)]
w = [x('rosen(x0)') for x in v]
w
[Sage, Sage, Sage]

Does anyone have any suggestion?

Sorry, I am new to this forum but already find it very helpful. I noticed that most posts have nicely formatted code snippets, but I couldn't figure out how to do it. When cut-and-paste from notebook() my code looks awful and unformatted.

Thanks for any suggestions.

In version 7.2 I no longer see the load balancing issue.

Istvan

I tried both and SAGENB_TMPDIR doesn't work, but setting TMPDIR does.

Thanks!

I am using Sage Notebook to generate very large files and then process them. The size of the files is tens of Gigabytes but that wouldn't be a problem the way things are set up in my worksheet. However, it seems no matter where the worksheet is stored the associated files are temporarily stored in /tmp whil ethe worksheet is working. For example,

$ ll /raid/istvan/Playground/Sage.sagenb/home/__store__/2/21/212/2123/admin/17/cells/4/
total 8
drwx------ 2 istvan istvan 4096 Sep  4 15:36 ./
drwxrwxr-x 9 istvan istvan 4096 Sep  4 15:34 ../
lrwxrwxrwx 1 istvan istvan   49 Sep  4 15:36 A.mmap -> /tmp/tmpa6CJmh/A.mmap
lrwxrwxrwx 1 istvan istvan   49 Sep  4 15:36 B.mmap -> /tmp/tmpa6CJmh/B.mmap
lrwxrwxrwx 1 istvan istvan   49 Sep  4 15:36 C.mmap -> /tmp/tmpa6CJmh/C.mmap
lrwxrwxrwx 1 istvan istvan   49 Sep  4 15:36 D.mmap -> /tmp/tmpa6CJmh/D.mmap

$ ll /tmp/tmpa6CJmh
total 68860
drwx------  2 istvan istvan        4096 Sep  4 15:35 ./
drwxrwxrwt 14 root   root         12288 Sep  4 15:35 ../
-rw-rw-r--  1 istvan istvan        1688 Sep  4 15:35 ___code___.py
-rw-rw-r--  1 istvan istvan 10000000000 Sep  4 15:37 A.mmap
lrwxrwxrwx  1 istvan istvan          54 Sep  4 15:35 data -> /raid/istvan/Playground/Sage.sagenb/home/admin/17/data/
-rw-rw-r--  1 istvan istvan 10000000000 Sep  4 15:37 B.mmap
-rw-rw-r--  1 istvan istvan 80000000000 Sep  4 15:37 C.mmap
-rw-rw-r--  1 istvan istvan 10000000000 Sep  4 15:37 D.mmap
-rw-rw-r--  1 istvan istvan        2219 Sep  4 15:35 _sage_input_6.py

The total size of the files is about 110 GB and I have plenty of room on the /raid partition where the Sage Notebook resides, but my / including /tmp partition is way too small for that. How can I make the Notebook NOT to write to /tmp? Is there an envar to set that?

Thanks for any suggestion,

Istvan

Download the package from http://www.hdfgroup.org/downloads/index.html and the installation is straightforward. On Linux:

1) untar the downloaded file
2) cd to the hdf5 directory
3) ./configure --prefix=/where/you/want/hdf5/to/be/installed (in my case it was /usr/local/hdf5)
4) make
5) make check
6) sudo make install (sudo needed if the location is not in your own user area)
7) sudo make check-install

I could answer my question as opposed to comment on it. It could be my browser, I don't know but let's consider this ticket closed.

Of course, hdf5 must be installed first and since it is not a Python package, pip will likely need explicit information about hdf5 libraries and include files. The following command worked for me:

$ sage -pip install --global-option=build_ext --global-option="-L/usr/local/hdf5/lib" --global-option="-l/usr/local/hdf5/lib" --global-option="-I/usr/local/hdf5/include" --global-option="-R/usr/local/hdf5/lib" h5py

I am trying to get further help regarding "Is HDF5 or the Python interface h5py supported in Sage?" I added my related question as a comment (while logged in) twice, the site seemed to have processed it but it won't show up under the above title. Should I open a new ticket?

UPDATE

This is fixed now :)

Got it. Great, thank you!

Istvan

I am saving very large list objects to disk using the save() command in Sage, which is utilizing Python's Pickles package. There is a known deficiency/bug that is unlikely to go away, namely that deeply buried in the compression code in the Python standard library that Pickles and therefore save() use, there are legacy 32-bit integers that result in a serious limitation in using save() with even moderately large (hundreds of MB) objects. See OverflowError: size does not fit in an int. The h5py package provides a Python interface to the HDF5 library and HDF5 can deal with multiple terabytes easily. Does anyone know if h5py is/will be implemented in Sage? Is there another alternative to save() in Sage to save objects to disk?

I am using v6.4.1.~/.sage/temp/ is different. If I run a Notebook session, its files are stored under ~/.sage/notebook/... but while a Sage calculation is running all the files there are just links to /tmp/something where /tmp/something is the value of SAGE_TMP in that session. I am saving large objects via the save() command and that's where I have the problem, because they can't fit in /tmp. There has got to be a way in Sage to set the prefix for SAGE_TMP from /tmp to something else.