% $OpenXM: OpenXM/doc/Papers/dagb-noro.tex,v 1.1 2001/10/03 08:32:58 noro Exp $
\setlength{\parskip}{10pt}
\begin{slide}{}
\fbox{A computer algebra system Risa/Asir}
\begin{itemize}
\item Old style software for polynomial computation
\begin{itemize}
\item Domain specification is not necessary prior to computation
\item automatic conversion of inputs into internal canonical forms
\end{itemize}
\item User language with C-like syntax
\begin{itemize}
\item No type declaration of variables
\item Builtin debugger for user programs
\end{itemize}
\item Open source
\begin{itemize}
\item Whole source tree is available via CVS
\end{itemize}
\item OpenXM interface
\begin{itemize}
\item As a client : can call procedures on other OpenXM servers
\item As a server : offers all its functionalities to OpenXM clients
\item As a library : OpenXM functionality is available via subroutine calls
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Major functionalities}
\begin{itemize}
\item Fundamental polynomial arithmetics
\begin{itemize}
\item Internal form of a polynomial : recursive representaion or distributed
representation
\end{itemize}
\item Polynomial factorization
\begin{itemize}
\item Univariate factorization over the rationals, algebraic number fields and various finite fields
\item Multivariate factorization over the rationals
\end{itemize}
\item Groebner basis computation
\begin{itemize}
\item Buchberger and $F_4$ algorithm
\item Change of ordering/RUR of 0-dimensional ideals
\item Primary ideal decomposition
\item Computation of $b$-function
\end{itemize}
\item PARI library interface
\item Paralell distributed computation under OpenXM
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{History of development : ---1994}
\begin{itemize}
\item --1989
Several subroutines were developed for a Prolog program.
\item 1989--1992
\begin{itemize}
\item Reconfigured as Risa/Asir with the parser and Boehm's conservative GC.
\item Developed univariate and multivariate factorizers over the rationals.
\end{itemize}
\item 1992--1994
\begin{itemize}
\item Started implementation of Groebner basis computation
User language $\Rightarrow$ rewritten in C (by Murao) $\Rightarrow$
trace lifting
\item Univariate factorization over algebraic number fields
Intensive use of successive extension, non-squarefree norms
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{History of development : 1994-1996}
\begin{itemize}
\item Free distribution of binary versions
\item Primary ideal decomposition
\begin{itemize}
\item Shimoyama-Yokoyama algorithm
\end{itemize}
\item Improvement of Buchberger algorithm
\begin{itemize}
\item Trace lifting+homogenization
\item Omitting check by compatible prime
\item Modular change of ordering, Modular RUR
\item Noro met Faug\`ere at RISC-Linz and he mentioned $F_4$.
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{History of development : 1996-1998}
\begin{itemize}
\item Distributed compuatation
\begin{itemize}
\item A prototype of OpenXM
\end{itemize}
\item Improvement of Buchberger algorithm
\begin{itemize}
\item Content reduction during nomal form computation
\item Its parallelization by the above facility
\item Application : computation of odd order replicable functions
Risa/Asir : it took 5days to compute a DRL basis ({\it McKay})
From Faug\`ere : computation of the DRL basis 53sec
\end{itemize}
\item Univariate factorization over large finite fields
\begin{itemize}
\item To implement Schoof-Elkies-Atkin algorithm
Counting rational points on elliptic curves --- not free
But related functions are freely available
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{History of development : 1998-2000}
\begin{itemize}
\item OpenXM
\begin{itemize}
\item OpenXM specification was written by Noro and Takayama
\item Functions for distributed computation were rewritten
\end{itemize}
\item Risa/Asir on Windows
\begin{itemize}
\item Requirement from a company for which Noro worked
Written in Visual C++
\end{itemize}
\item Test implementation of $F_4$
\begin{itemize}
\item Over $GF(p)$ : pretty good
\item Over the rationals : not so good except for {\it McKay}
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{History of development : 2000-current}
\begin{itemize}
\item The source code is freely available
\begin{itemize}
\item Noro moved from Fujitsu to Kobe university.
\item Fujitsu kindly permitted to make Risa/Asir open source.
\end{itemize}
\item OpenXM
\begin{itemize}
\item Revising the specification : OX-RFC100, 101, (102)
\item OX-RFC102 : ommunications between servers via MPI
\end{itemize}
\item Rings of differential operators
\begin{itemize}
\item Buchberger algorithm
\item $b$-function computation
Minimal polynomial computation by modular method
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Status of each component --- Factorizer}
\begin{itemize}
\item 10 years ago
its performace was fine compared with existing software
like REDUCE, Maple, Mathematica.
\item 4 years ago
Univarate factorization over algebraic number fields was
still fine because of some tricks on factoring polynomials
derived from norms.
\item Current
Multivariate : not so bad
Univariate : completely obsolete by M. van Hoeij's new algorithm
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Status of each component --- Groebner basis related functions}
\begin{itemize}
\item 8 years ago
The performace was poor with only the sugar strategy.
\item 7 years ago
Rather fine with trace lifting but Faug\`ere's (old)Gb was more
efficient.
Homogenization+trace lifting made it possible to compute
wider range of Groebner bases.
\item 4 years ago
Modular RUR was comparable with Rouillier's implementation.
\item Current
FGb seems much more efficient than our $F_4$ implementation.
Singular's Groebner basis computation is also several times
faster than Risa/Asir, because Singular seems to have efficient
monomial and polynomial representation.
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Ground fields}
\begin{itemize}
\item The rational number field
\item Algebraic number fields
represented by successive extensions
\item Finite fields
\begin{itemize}
\item $GF(p)$ ($p<2^{30}$); represented by a single word
\item $GF(p^n)$ ($p^n < 2^{20}$); represented by a primitive root
\item $GF(2^n)$; represented by a bit string
\item $GF(p)$ ($p$ : arbitrary prime)
\item $GF(p^n)$ ($p$ : arbitrary odd prime)
\end{itemize}
\item Real and complex number fields
\begin{itemize}
\item Double float
\item PARI bigfloat
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Polynomial factorization}
\begin{itemize}
\item Univariate factorization
\begin{itemize}
\item Over the rationals
Berlekamp-Zassenhaus
(classical; knapsack has not yet implemented)
\item Over algebraic number fields
Trager's algorithm + some improvement
\item Over finite fieds
DDF + Cantor-Zassenhaus; FFT for large finite fields
\end{itemize}
\item Multivariate factorization
\begin{itemize}
\item Over the rationals
Classical EZ algorithm
\item Over finite fieds
Modified Bernardin square free, bivariate Hensel
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Groebner basis computation --- Buchberger algorithm}
\begin{itemize}
\item Polynomial rings
\begin{itemize}
\item Over finite fields
Any finite field is available as a ground field
\item Over the rationals
Guess of a groebner basis by detecting zero reduction by modular computation
+ check (trace lifting)
Homogenization+guess+dehomogenization+check
\end{itemize}
\item Rings of differential operators
\begin{itemize}
\item Groebner basis of a left ideal
An efficient implementation of Leibniz rule
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{$F_4$ algorithm}
\begin{itemize}
\item Over small finite fields ($GF(p)$, $p < 2^{30}$)
\begin{itemize}
\item More efficient than Buchberger algorithm
but less efficient than FGb by Faugere
\end{itemize}
\item Over the rationals
\begin{itemize}
\item Very naive implementation
\item Less efficient than Buchberger algorithm
except for one example
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Change of ordering for zero-dimimensional ideals}
\begin{itemize}
\item Any ordering to lex ordering
\begin{itemize}
\item Modular change of ordering
Guess of the support by modular FGLM
+ solving linear systems by Hensel lifting
\end{itemize}
\item RUR (generalized shape lemma)
\begin{itemize}
\item Modular RUR (only implemented on the shape base case)
Almost the same as modular change of ordering
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Primary decomposition --- Shimoyama-Yokoyama algorithm}
\begin{itemize}
\item Only implemented over the rationals
Finite field version will soon be available
\item Pseudo primary ideal
An ideal whose radical is prime
\item Prime decomp. of the radical $\Rightarrow$ pseudo primary decomp.
\item Extraction of embedded components
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Computation of $b$-function}
$D$ : the ring of differential operators
$b(s)$ : a polynomial of the smallest degree s.t.
there exists $P(s) \in D[s]$, $P(s)f^{s+1}=b(s)f^s$
$b(s)$ : $b$-function of a polynomial $f$
$\Rightarrow$ $b(s)$ can be computed by Oaku's algorithm
On Risa/Asir : $b(s)$ is computed by
A groebner basis $\Rightarrow$ guess of $\deg(b(s))$ by modular
computation $\Rightarrow$ solving a linear system
\end{slide}
\begin{slide}{}
\fbox{Interface to PARI library}
\begin{itemize}
\item Data conversion
\begin{itemize}
\item Only primitive data can be passed to PARI
Rational number, bignum, bigfloat, polynomial,
vector, matrix
\item Results are converted to Risa objects
\end{itemize}
\item Evaluation of transcendental functions
\begin{itemize}
\item Most of univariate functions in PARI can be
evaluated by {\tt eval()}
\end{itemize}
\item Calling PARI functions
\begin{itemize}
\item One can call PARI functions by {\tt pari({\it parifunction},{\it args})}
The knapsack factorization is available via {\tt pari(factor,{\it poly})}
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{OpenXM}
\begin{itemize}
\item An environment for parallel distributed computation
Both for interactive, non-interactive environment
\item Message passing
OX (OpenXM) message : command and data
\item Hybrid command execution
\begin{itemize}
\item Stack machine command
push, pop, function execution, $\ldots$
\item accepts its own command sequences
{\tt execute\_string} --- easy to use
\end{itemize}
\item Data is represented as CMO
CMO --- Common Mathematical Object format
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{OpenXM server interface in Risa/Asir}
\begin{itemize}
\item TCP/IP stream
\begin{itemize}
\item Launcher
A client executes a launcher on a host.
The launcher launches a server on the same host.
\item Server
A server reads from the descriptor 3, write to the descriptor 4.
\end{itemize}
\item Subroutine call
Risa/Asir subroutine library provides interfaces corresponding to
pushing and popping data and executing stack commands.
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{OpenXM client interface in Risa/Asir}
\begin{itemize}
\item Primitive interface functions
Pushing and popping data, sending commands etc.
\item Convenient functions
Launching servers, calling remote functions,
interrupting remote executions etc.
\item Parallel distributed computation is easy
Simple parallelization is practically important
Competitive computation is easily realized
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{CMO = Serialized representation of mathematical object}
\begin{itemize}
\item primitive data
\begin{eqnarray*}
\mbox{Integer32} &:& ({\tt CMO\_INT32}, {\sl int32}\ \mbox{n}) \\
\mbox{Cstring}&:& ({\tt CMO\_STRING},{\sl int32}\, \mbox{ n}, {\sl string}\, \mbox{s}) \\
\mbox{List} &:& ({\tt CMO\_LIST}, {\sl int32}\, len, ob[0], \ldots,ob[m-1])
\end{eqnarray*}
\item numbers and polynomials
\begin{eqnarray*}
\mbox{ZZ} &:& ({\tt CMO\_ZZ},{\sl int32}\, {\rm f}, {\sl byte}\, \mbox{a[1]}, \ldots
{\sl byte}\, \mbox{a[$|$f$|$]} ) \\
\mbox{Monomial32}&:& ({\tt CMO\_MONOMIAL32}, n, \mbox{e[1]}, \ldots, \mbox{e[n]}, \mbox{Coef}) \\
\mbox{Coef}&:& \mbox{ZZ} | \mbox{Integer32} \\
\mbox{Dpolynomial}&:& ({\tt CMO\_DISTRIBUTED\_POLYNOMIAL},\\
& & m, \mbox{DringDefinition}, \mbox{Monomial32}, \ldots)\\
\mbox{DringDefinition}
&:& \mbox{DMS of N variables} \\
& & ({\tt CMO\_RING\_BY\_NAME}, name) \\
& & ({\tt CMO\_DMS\_GENERIC}) \\
\end{eqnarray*}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Stack based communication}
\begin{itemize}
\item Data arrived a client
Pushed to the stack
\item Result
Pushd to the stack
Written to the stream when requested by a command
\item The reason why we use the stack
\begin{itemize}
\item Stack = I/O buffer for (possibly large) objects
Multiple requests can be sent before their exection
A server does not get stuck in sending results
\end{itemize}
\end{itemize}
\end{slide}
\begin{slide}{}
\fbox{Executing functions on a server (I) --- {\tt SM\_executeFunction}}
\begin{enumerate}
\item (C $\rightarrow$ S) Arguments are sent in binary encoded form.
\item (C $\rightarrow$ S) The number of aruments is sent as {\sl Integer32}.
\item (C $\rightarrow$ S) A function name is sent as {\sl Cstring}.
\item (C $\rightarrow$ S) A command {\tt SM\_executeFunction} is sent.
\item The result is pushed to the stack.
\item (C $\rightarrow$ S) A command {\tt SM\_popCMO} is sent.
\item (S $\rightarrow$ C) The result is sent in binary encoded form.
\end{enumerate}
$\Rightarrow$ Communication is fast, but functions for binary data
conversion are necessary.
\end{slide}
\begin{slide}{}
\fbox{Executing functions on a server (II) --- {\tt SM\_executeString}}
\begin{enumerate}
\item (C $\rightarrow$ S) A character string represeting a request in a server's
user language is sent as {\sl Cstring}.
\item (C $\rightarrow$ S) A command {\tt SM\_executeString} is sent.
\item The result is pushed to the stack.
\item (C $\rightarrow$ S) A command {\tt SM\_popString} is sent.
\item (S $\rightarrow$ C) The result is sent in readable form.
\end{enumerate}
$\Rightarrow$ Communication may be slow, but the client parser may be
enough to read the result.
\end{slide}
\begin{slide}{}
\fbox{Example of distributed computation --- $F_4$ vs. $Buchberger$ }
\begin{verbatim}
/* competitive Gbase computation over GF(M) */
/* Cf. A.28 in SINGULAR Manual */
/* Process list is specified as an option : grvsf4(...|proc=P) */
def grvsf4(G,V,M,O)
{
P = getopt(proc);
if ( type(P) == -1 ) return dp_f4_mod_main(G,V,M,O);
P0 = P[0]; P1 = P[1]; P = [P0,P1];
map(ox_reset,P);
ox_cmo_rpc(P0,"dp_f4_mod_main",G,V,M,O);
ox_cmo_rpc(P1,"dp_gr_mod_main",G,V,0,M,O);
map(ox_push_cmd,P,262); /* 262 = OX_popCMO */
F = ox_select(P); R = ox_get(F[0]);
if ( F[0] == P0 ) { Win = "F4"; Lose = P1;}
else { Win = "Buchberger"; Lose = P0; }
ox_reset(Lose); /* simply resets the loser */
return [Win,R];
}
\end{verbatim}
\end{slide}
\begin{slide}{}
\end{slide}
\begin{slide}{}
\end{slide}
\begin{slide}{}
\end{slide}
\begin{slide}{}
\end{slide}
\begin{slide}{}
\end{slide}