Karim Belabas on Fri, 04 Feb 2005 19:44:58 +0100 |
[Date Prev] [Date Next] [Thread Prev] [Thread Next] [Date Index] [Thread Index]
Re: infinite loop using qflllgram on an opteron |
* Lars Fischer [2005-01-20 22:58]: > I am running into trouble using qflllgram on an Opteron machine, while > the same calculation on 32-bit AMD-XP and Athlon-64 works without problems. > > > I have several thousand 70x70 matrices G_i with integer entries. Each > entry is about 100 digits long. > > On an Opteron with pari-2.2.9: > qflllgram(G_i) takes about 2 minutes for 95 percent of the i's . > But in 5 percent of my matrices, the calculation runs for hours, without > any result. > > Now I tried several things with some of the problem matrices: > -On a 32 bit Desktop AMD-XP and a Athlon-64 the calculation finished > without Problem. > -I compiled pari without optimizations -O0, no change. > -Using pari-2.2.8 on the Opteron, some of my problem matrices finished > also, but the rest did not. > > I found > http://pari.math.u-bordeaux.fr/archives/pari-dev-0301/msg00048.html > It is suggested to increase the memory. I tried it with 130, 230, 330, > 500, 800 and 1000 MB. Always the same: an infinite loop. (And for 95 > percent of the matrices 130 MB was enough.) > > I inserted \gm 2, as also suggested. The output ends like You should also add \g4. It would show a large number of precision doubling. > Has anybody noticed the same problem? > > Can you give me advice how to avoid it? You can always use qflllgram(x, 1) which will be much slower [ generically ] but is guaranteed to converge. Can you send me a problematic matrix ? I'll try and see whether the Opteron (or compiler) is at fault, or my strategy. [ What's the compiler by the way ? ] Thanks! Karim. PS: a new adaptive Schnorr-Euchner scheme for LLL-reduction [with guaranteed convergence, and still good theoretical complexity] has recently been proposed by NGuyen & Stehle. It is pending implementation in PARI... -- Karim Belabas Tel: (+33) (0)1 69 15 57 48 Dep. de Mathematiques, Bat. 425 Fax: (+33) (0)1 69 15 60 19 Universite Paris-Sud http://www.math.u-psud.fr/~belabas/ F-91405 Orsay (France) http://pari.math.u-bordeaux.fr/ [PARI/GP]