Opened 12 years ago
Closed 11 years ago
#4237 closed enhancement (fixed)
[with patch; positive review] magma -- finite field matrix conversions
Reported by: | was | Owned by: | was |
---|---|---|---|
Priority: | major | Milestone: | sage-3.2.2 |
Component: | interfaces | Keywords: | |
Cc: | Merged in: | ||
Authors: | Reviewers: | ||
Report Upstream: | Work issues: | ||
Branch: | Commit: | ||
Dependencies: | Stopgaps: |
Description
2) Converting matrices (over finite fields) is very slow for the dimensions I'm interested in (the smallest dimension is 10000x10000) because it each element is converted individually. Also, the conversion eats a lot of RAM due to the large string that is created.
Reported by Martin Albrecht
Attachments (3)
Change History (23)
comment:1 Changed 12 years ago by
- Status changed from new to assigned
comment:2 Changed 11 years ago by
comment:3 Changed 11 years ago by
Which version of magma are you using and on which operating system? I think there is a major bug in Magma-2.14 on OS X where it takes a vast amount of memory to parse any input above a certain threshhold (that threshhold is maybe 1000x1000 matrices). Doing the same computation on sage.math (Linux) and it takes almost no memory. There's not much I can do about that, except complain to Magma.
William
comment:4 Changed 11 years ago by
I think something can be done:
sage: A = random_matrix(GF(2),2*10^4,2*10^4) sage: AM = magma(A)
This will create a *huge* string in memory (in the Sage process), write that to disk and then read the file into Magma. We could at least write the strings in smaller chunks to disk to avoid the RAM consumption on the Sage side.
Btw. Magma crashes on me on sage.math:
before (last 100 chars): home/was/s/data/extcode//magma/latex/latex.m is compiled for newer version than current executable
comment:5 Changed 11 years ago by
Btw. Magma crashes on me on sage.math:
That's because I installed a newer magma for myself locally, and it overwrites the .sig files, so the system-wide magma doesn't work. I deleted them and recreated them with the old version. If you use your own copy of sage this won't be an issue.
comment:6 Changed 11 years ago by
Regarding your example
sage: A = random_matrix(GF(2),2*10^4,2*10^4)
I don't think there is any way to reasonably expect to use pexpect or a file to convert that matrix to Magma quickly using anything like is currently done in Sage or like you suggest even. Just doing A.str() takes nearly 2GB and takes about the same amount of RAM as A._magma_init_(magma):
sage: A = random_matrix(GF(2),2*10^4,2*10^4) sage: time s = A._magma_init_(magma) pCPU times: user 203.00 s, sys: 7.73 s, total: 210.73 s Wall time: 211.47 s sage: get_memory_usage() 1274.81640625 sage: len(s) 800000011 Versus sage: A = random_matrix(GF(2),2*10^4,2*10^4) sage: b = str(A) sage: b = A.str() sage: get_memory_usage() 1273.48828125 sage: len(b) 800039999
Writing to a file as we go would reduce memory usage at any given moment, but will take as long or longer -- which means several minutes -- which seems unacceptable.
The optimal solution for your application would be to write a Sage function that writes to file a mod-2 matrix in a single packed binary format, and a Magma function that reads that file and makes a matrix. As far as I can tell, the *ONLY* way to turn a string into anything of any use in Magma is to use the function StringToInteger?, which fortunately has a version that takes a base:
> StringToInteger("af",16); 175
In all cases, it would be immensely useful for this if there were a Sage function like this:
sage: A = random_matrix(GF(2),2*10^4,2*10^4) sage: n = A._as_packed_integer()
The output would be an integer that has 2*10^4*2*10^4
digits in base 2, or
2*10^4*2*10^4/4 = 100000000 = 10^8
digits in base 16. Computing the base-16 string representation of such a massive number in Sage takes about a half second:
sage: b = 2*10^4*2*10^4 sage: s = ZZ.random_element(x=0,y=2^b) sage: time len(s.str(16)) CPU times: user 0.45 s, sys: 0.18 s, total: 0.63 s
Memory usage for storing the number and *also* the string representation is about 111MB, which is reasonable (much better than several GB, which is the current situation).
Writing our massive base-16 string to a file takes about 0.5 seconds.
sage: time open('foo','w').write(t) CPU times: user 0.00 s, sys: 0.56 s, total: 0.56 s Wall time: 0.56 s
Reading it into Magma takes about 2.5 seconds:
sage: magma.eval('time A := Read("foo");') 'Time: 2.260'
Incidentally, Python is literally ALMOST TEN TIMES faster at reading in a file that Magma!!
sage: time k = open('foo').read() CPU times: user 0.00 s, sys: 0.28 s, total: 0.28 s
Anyway, now that we have that string in Magma, we convert it to an integer, which takes 3.36 seconds:
sage: magma.eval('time n := StringToInteger(A, 16);') 'Time: 3.360'
Incidentally, the same operation in Sage is over twice as faster:
sage: time k = ZZ(t, base=16) CPU times: user 1.14 s, sys: 0.25 s, total: 1.39 s Wall time: 1.39 s
OK, so the one remaining step is to convert that big integer in Magma into a boolean matrix in Magma. This is probably going to be the killer bottleneck. ... In fact, I'm totally stumped about how to turn a packed integer that corresponds to a matrix over GF(2) into anything useful. There seems to be no "bitwise or" on integers, no way to get the ith bit of an integer, nothing! All I can think to do is "mod 2" and divide by 2 repeatedly, but dividing by 2 is very expensive.
There is the command StringToIntegerSequence?, but it stupidly only works with numbers in base 10, making it utterly and completely useless.
I'm stumped.
comment:7 Changed 11 years ago by
For the record using StringToIntegerSequence
on 0/1 integers, I can transfer 4*10^{8 zeros and ones from Sage to Magma in 102 seconds, where the time is totally dominated by Magma's StringToIntegerSequence? command.
}
sage: time open('/home/was/foo','w').write('1 '*10^8) CPU time: 1.58 s, Wall time: 1.58 s sage: %magma sage: time a := Read("/home/was/foo"); sage: time b := StringToIntegerSequence(a); Time: 4.180 Time: 18.910 sage: time open('/home/was/foo','w').write('1 '*(2*10^4*2*10^4)) CPU time: 6.38 s, Wall time: 6.37 s sage: %magma sage: time a := Read("/home/was/foo"); sage: time b := StringToIntegerSequence(a); Time: 17.410 Time: 79.150 sage: 17.4 + 79.15 + 6.37 102.920000000000
Incidentally, I estimate that it would take 326 seconds (isntead of 79 seconds) to do the same thing Sage's StringToIntegerSequence
does in pure Python:
sage: v = '1 '*(2*10^3*2*10^3) sage: time z =[int(a) for a in v.split()] CPU time: 3.13 s, Wall time: 3.26 s
One would have to use Cython or something to beat that, though it doesn't matter for what we're doing.
Back to our story. So if we use the Magma parser itself to read in a boolean vector of length 4*10^{8, I estimate it will take 384 seconds just for reading, which is about 4 times as long as above, so StringToIntegerSequence? beats using the parser. }
sage: n = 10^7 sage: s = ('v := Vector(GF(2),%s,%s);'%(n, [1]*n)).replace(' ', '') sage: time open('/home/was/foo','w').write(s) Time: CPU 0.12 s, Wall: 0.13 s sage: time magma.eval('load "/home/was/foo";') 'Loading "/home/was/foo"' CPU time: 0.00 s, Wall time: 9.60 s sage: n = 2*10^7 sage: s = ('v := Vector(GF(2),%s,%s);'%(n, [1]*n)).replace(' ', '') sage: time open('/home/was/foo','w').write(s) Time: CPU 0.23 s, Wall: 0.23 s sage: time magma.eval('load "/home/was/foo";') 'Loading "/home/was/foo"' CPU time: 0.00 s, Wall time: 18.72 s sage: 4*10^8/10^7 * 9.6 384.000000000000
Interestingly, just using the Magma parser to read in a sequence of 1's in decimal without putting Vector(GF(2), n, ...)
around the sequence takes WAY longer, i.e., I estimate it would take about 40,000 seconds (!):
sage: n = 10^5 sage: time s = ('v := %s;'%[1]*n).replace(' ', '') sage: time open('/home/was/foo','w').write(s) Time: CPU 0.02 s, Wall: 0.02 s Time: CPU 0.01 s, Wall: 0.01 s sage: time magma.eval('load "/home/was/foo";') 'Loading "/home/was/foo"' CPU time: 0.00 s, Wall time: 10.08 s sage: 4*10^8/10^5 * 10.08 40320.0000000000
I tried the above several times -- it is just shockingly slow.
Some other weirdness:
- It takes magma 123 times as long to parse the hex integer literal version of
3^(10^6)
than the base-10 integer literal version of that number!!! Also, using StringToInteger? takes over 100 times longer than just using the load command on an assignment. I'm trying a bunch of different variations... and the winner is:21:08 < wstein> If you do "v := 0x<big hex string>;" in Magma it is insanely slow. 21:08 < wstein> If you do "v := <big decimal string>;" it is fast (like sage) 21:08 < wstein> If you do "v := StringToInteger(<big decimal string>)" it is insansely slow. 21:09 < wstein> If you do "v := StringToInteger(0x<big hex string>, 16)" it is very fast, just like sage. 21:09 < wstein> So the last option is good :)
comment:8 Changed 11 years ago by
The first attached patch makes conversion to Magma for matrix_modn_dense twice as fast as before, and somewhat more memory efficient.
BEFORE:
age: w = random_matrix(GF(97),2000) sage: time m = magma(w) CPU times: user 3.00 s, sys: 0.46 s, total: 3.47 s Wall time: 7.46 s
AFTER:
sage: w = random_matrix(GF(97),2000) sage: time m = magma(w) CPU times: user 1.08 s, sys: 0.19 s, total: 1.27 s Wall time: 3.22 s
This does not apply to matrices over GF(2), which still require more special code to be faster.
This patch also fixes a free that should be a sage_free, and speeds up the list method for matrices mod n.
comment:9 Changed 11 years ago by
- Summary changed from magma -- finite field matrix conversions to [with patch; half done] magma -- finite field matrix conversions
comment:10 Changed 11 years ago by
trac_4237_part1.patch
assumes that we always store the entries as positive integers. Is this a sensible assumption?
comment:11 Changed 11 years ago by
trac_4237_part1.patch assumes that we always store the entries as positive integers. Is this a sensible assumption?
Yes. All entries are stored as numbers between 0 and p-1. See, e.g., code like this in the init method:
if PY_TYPE_CHECK_EXACT(x, int): tmp = (<long>x) % p v[j] = tmp + (tmp<0)*p
William
comment:12 Changed 11 years ago by
- Summary changed from [with patch; half done] magma -- finite field matrix conversions to [with patch; needs review] magma -- finite field matrix conversions
The second patch adds support for much faster conversions of matrices over GF(2) to Magma. Now sage can convert your 10000x10000 GF(2) matrix in about 45 seconds, instead of the 149 seconds it used to take (so over 3 times faster). The intermediate memory usage is also better on the Sage side.
BEFORE:
sage: w = random_matrix(GF(2),10000) sage: time k = w._magma_init_(magma) CPU times: user 49.64 s, sys: 2.20 s, total: 51.84 s Wall time: 52.75 s sage: time a = magma(k) CPU times: user 1.94 s, sys: 2.69 s, total: 4.63 s Wall time: 97.84 s sage: 52 + 97 149
AFTER:
sage: w = random_matrix(GF(2),10000) sage: time k = w._magma_init_(magma) CPU times: user 1.22 s, sys: 0.63 s, total: 1.85 s Wall time: 1.85 s sage: time a = magma(k) CPU times: user 1.70 s, sys: 2.30 s, total: 4.00 s Wall time: 44.15 s
Now the time is totally dominated by the time Magma spends in the function StringToIntegerSequence?.
I tried the 2*10^{4 example. With the new code, Sage does create a 5GB string in memory, then writes it to a 763MB file: }
was@sage:~/.sage/temp/sage/22977/interface$ ls -lh total 764M -rw-r--r-- 1 was was 763M 2008-12-12 11:12 tmp22977
Then Sage frees all that memory. Then Magma reads in the string to memory, which takes over 5GB again (not surprisingly):
23012 was 25 0 5772m 5.1g 2796 R 100 8.0 2:41.25 magma.exe.x86_6
Converts that to a string over integers (which also takes a lot of RAM), then creates a mod-2 matrix.
It would in theory be possible to do this via a series of chunks as you suggest, but support for that does not exist in the Magma/Sage? interface (or any interface) at all, and would probably be particular hard to get right when Magma isn't even running locally (I think now one could have the magma run on a different machine and that 763MB file above would get copied via scp, maybe...)
If only Magma had any support WHATEVER for bit manipulation. Then we could send the whole data over as a single integer or other format and extract it. But Magma doesn't. At present it seems that by far the fastest way to get a mod-2 matrix over to Magma is to do what I've implemented in patch 2, which is pretty bad memory-wise in the limit. It does complete your 2*10^{4 example on sage.math in 192 seconds, which is an improvement: }
sage: A = random_matrix(GF(2),2*10^4,2*10^4) sage: time B = magma(A) CPU times: user 16.28 s, sys: 14.38 s, total: 30.65 s Wall time: 192.45 s
Anyway, for something more memory efficient, what I just wrote could be the base case. Maybe you could try sending a big matrix over GF(2) to magma by breaking it into smaller matrices, sending each of those (with the new code I just wrote), then reassembling the result in Magma. Can magma do things like stack matrices (like Sage's A.stack)?
Changed 11 years ago by
comment:13 Changed 11 years ago by
patch2 doesn't apply against my 3.2.1. Hunk:
--- matrix_mod2_dense.pyx +++ matrix_mod2_dense.pyx @@ -1136,7 +1136,7 @@ EXAMPLE: sage: A = random_matrix(GF(2),3,3) sage: A._magma_init_(magma) # optional - magma - '_sage_[...]![0,1,0,0,1,1,0,0,0]' + 'Matrix(GF(2),3,3,StringToIntegerSequence("0 1 0 0 1 1 0 0 0"))' sage: A = random_matrix(GF(2),100,100) sage: B = random_matrix(GF(2),100,100) sage: magma(A*B) == magma(A) * magma(B) # optional - magma
Do I need 3.2.2.alphaX for this patch?
comment:14 Changed 11 years ago by
- Summary changed from [with patch; needs review] magma -- finite field matrix conversions to [with patch; needs work] magma -- finite field matrix conversions
I am seeing one doctest failure with this patch on sage.math:
sage -t -long "devel/sage/sage/matrix/matrix_modn_dense.pyx" ********************************************************************** File "/scratch/mabshoff/release-cycle/sage-3.2.2.rc0/devel/sage/sage/matrix/matrix_modn_dense.pyx", line 276, in __main__.example_6 Failed example: -m###line 554:_sage_ >>> -m Expected: [19 18 17] [16 15 14] [13 12 11] Got: [ 0 18 17] [16 15 14] [13 12 11] **********************************************************************
I haven't looked into the cause yet.
Cheers,
Michael
comment:15 Changed 11 years ago by
A quick comment. The *new* output you're getting is *right*. The old output, which is in sage, is wrong. Somebody (probably me) fixed a bug. The matrix is a matrix modulo 19, so its entries should be normalized between 0 and 18, inclusive. That 19 should be 0, as it is in the output you get now.
comment:16 Changed 11 years ago by
- Summary changed from [with patch; needs work] magma -- finite field matrix conversions to [with patch; needs review] magma -- finite field matrix conversions
The doctest that "fails" (by giving correct output) is *not* in the patch I submitted. In my code I fixed some free's that should be sage_free, and implemented a correct specialized faster list method. Probably one of those two fixed the bug.
comment:17 Changed 11 years ago by
OK, I fixed the doctest, then fixed the neg method which had a bug (and has ever since it was written). This has nothing to do with the sage/magma interface, by the way. The doctest used to be wrong because of that bug; the new list method I wrote fixed the bug in printing, but didn't fix the actual bug. Now it's fixed. The bug was in negation of a number modulo p and forgetting to normalize the result in the range 0 to p-1, inclusive.
Changed 11 years ago by
comment:18 Changed 11 years ago by
- Milestone changed from sage-3.4 to sage-3.2.2
comment:19 Changed 11 years ago by
- Summary changed from [with patch; needs review] magma -- finite field matrix conversions to [with patch; positive review] magma -- finite field matrix conversions
Positive review for all three patches.
Cheers,
Michael
comment:20 Changed 11 years ago by
- Resolution set to fixed
- Status changed from assigned to closed
Merged all three patches in Sage 3.2.2.rc1
I am particularly interrested in GF(2) but other finite fields should behave the same. To reproduce: