Annotation of OpenXM/doc/ascm2001/homogeneous-network.tex, Revision 1.3
1.3 ! noro 1: % $OpenXM: OpenXM/doc/ascm2001/homogeneous-network.tex,v 1.2 2001/03/07 07:17:02 noro Exp $
1.1 noro 2:
3: \subsection{Distributed computation with homogeneous servers}
4: \label{section:homog}
5:
6: One of the aims of OpenXM is a parallel speedup by a distributed computation
7: with homogeneous servers. As the current specification of OpenXM does
8: not include communication between servers, one cannot expect
9: the maximal parallel speedup. However it is possible to execute
10: several types of distributed computation as follows.
11:
12: \subsubsection{Product of univariate polynomials}
13:
14: Shoup \cite{Shoup} showed that the product of univariate polynomials
15: with large degrees and large coefficients can be computed efficiently
16: by FFT over small finite fields and Chinese remainder theorem.
17: It can be easily parallelized:
18:
19: \begin{tabbing}
20: Input :\= $f_1, f_2 \in {\bf Z}[x]$ such that $deg(f_1), deg(f_2) < 2^M$\\
21: Output : $f = f_1f_2$ \\
22: $P \leftarrow$ \= $\{m_1,\cdots,m_N\}$ where $m_i$ is an odd prime, \\
23: \> $2^{M+1}|m_i-1$ and $m=\prod m_i $ is sufficiently large. \\
24: Separate $P$ into disjoint subsets $P_1, \cdots, P_L$.\\
25: for \= $j=1$ to $L$ $M_j \leftarrow \prod_{m_i\in P_j} m_i$\\
26: Compute $F_j$ such that $F_j \equiv f_1f_2 \bmod M_j$\\
27: \> and $F_j \equiv 0 \bmod m/M_j$ in parallel.\\
28: \> (The product is computed by FFT.)\\
29: return $\phi_m(\sum F_j)$\\
30: (For $a \in {\bf Z}$, $\phi_m(a) \in (-m/2,m/2)$ and $\phi_m(a)\equiv a \bmod m$)
31: \end{tabbing}
32:
33: Figure \ref{speedup}
34: shows the speedup factor under the above distributed computation
35: on Risa/Asir. For each $n$, two polynomials of degree $n$
36: with 3000bit coefficients are generated and the product is computed.
37: The machine is FUJITSU AP3000,
38: a cluster of Sun workstations connected with a high speed network
39: and MPI over the network is used to implement OpenXM.
40: \begin{figure}[htbp]
41: \epsfxsize=8.5cm
42: \epsffile{speedup.ps}
43: \caption{Speedup factor}
44: \label{speedup}
45: \end{figure}
46:
47: If the number of servers is $L$ and the inputs are fixed, then the cost to
48: compute $F_j$ in parallel is $O(1/L)$, whereas the cost
49: to send and receive polynomials is $O(L)$ if {\tt ox\_push\_cmo()} and
50: {\tt ox\_pop\_cmo()} are repeatedly applied on the client.
51: Therefore the speedup is limited and the upper bound of
52: the speedup factor depends on the ratio of
53: the computational cost and the communication cost for each unit operation.
54: Figure \ref{speedup} shows that
55: the speedup is satisfactory if the degree is large and $L$
56: is not large, say, up to 10 under the above environment.
1.2 noro 57: If OpenXM provides collective operations for broadcast and reduction
1.1 noro 58: such as {\tt MPI\_Bcast} and {\tt MPI\_Reduce} respectively, the cost of
59: sending $f_1$, $f_2$ and gathering $F_j$ may be reduced to $O(\log_2L)$
1.2 noro 60: and we can expect better results in such a case. In order to implement
61: such operations we need new specifications for inter-sever communication
62: and the session management. The will be proposed as OpenXM-RFC-102 in future.
63: We note that preliminary experiments shows the collective operations
64: works well on OpenXM.
1.1 noro 65:
66: \subsubsection{Competitive distributed computation by various strategies}
67:
68: SINGULAR \cite{Singular} implements {\it MP} interface for distributed
69: computation and a competitive Gr\"obner basis computation is
70: illustrated as an example of distributed computation.
71: Such a distributed computation is also possible on OpenXM.
72: The following Risa/Asir function computes a Gr\"obner basis by
73: starting the computations simultaneously from the homogenized input and
74: the input itself. The client watches the streams by {\tt ox\_select()}
75: and the result which is returned first is taken. Then the remaining
76: server is reset.
77:
78: \begin{verbatim}
79: /* G:set of polys; V:list of variables */
80: /* O:type of order; P0,P1: id's of servers */
81: def dgr(G,V,O,P0,P1)
82: {
83: P = [P0,P1]; /* server list */
84: map(ox_reset,P); /* reset servers */
85: /* P0 executes non-homogenized computation */
86: ox_cmo_rpc(P0,"dp_gr_main",G,V,0,1,O);
87: /* P1 executes homogenized computation */
88: ox_cmo_rpc(P1,"dp_gr_main",G,V,1,1,O);
89: map(ox_push_cmd,P,262); /* 262 = OX_popCMO */
90: F = ox_select(P); /* wait for data */
91: /* F[0] is a server's id which is ready */
92: R = ox_get(F[0]);
93: if ( F[0] == P0 ) {
94: Win = "nonhomo"; Lose = P1;
95: } else {
96: Win = "homo"; Lose = P0;
97: }
98: ox_reset(Lose); /* reset the loser */
99: return [Win,R];
100: }
101: \end{verbatim}
1.2 noro 102:
103: \subsubsection{Nesting of client-server communication}
104:
105: Under OpenXM-RFC-100 an OpenXM server can be a client of other servers.
106: Figure \ref{tree} illustrates a tree-like structure of an OpenXM
107: client-server communication.
108: \begin{figure}
109: \label{tree}
110: \begin{center}
111: \begin{picture}(200,140)(0,0)
112: \put(70,120){\framebox(40,15){client}}
113: \put(20,60){\framebox(40,15){server}}
114: \put(70,60){\framebox(40,15){server}}
115: \put(120,60){\framebox(40,15){server}}
116: \put(0,0){\framebox(40,15){server}}
117: \put(50,0){\framebox(40,15){server}}
118: \put(135,0){\framebox(40,15){server}}
119:
120: \put(90,120){\vector(-1,-1){43}}
121: \put(90,120){\vector(0,-1){43}}
122: \put(90,120){\vector(1,-1){43}}
123: \put(40,60){\vector(-1,-2){22}}
124: \put(40,60){\vector(1,-2){22}}
125: \put(140,60){\vector(1,-3){14}}
126: \end{picture}
127: \caption{Tree-like structure of client-server communication}
128: \end{center}
129: \end{figure}
130: Such a computational model is useful for parallel implementation of
131: algorithms whose task can be divided into subtasks recursively. A
132: typical example is {\it quicksort}, where an array to be sorted is
133: partitioned into two sub-arrays and the algorithm is applied to each
134: sub-array. In each level of recursion, two subtasks are generated
135: and one can ask other OpenXM servers to execute them. Though
1.3 ! noro 136: this makes little contribution to the efficiency, it is worth
1.2 noro 137: to show that such an attempt is very easy under OpenXM.
138: Here is an Asir program.
139: A predefined constant {\tt LevelMax} determines
140: whether new servers are launched or whole subtasks are done on the server.
141:
142: \begin{verbatim}
143: #define LevelMax 2
144: extern Proc1, Proc2;
145: Proc1 = -1$ Proc2 = -1$
146:
147: /* sort [A[P],...,A[Q]] by quicksort */
148: def quickSort(A,P,Q,Level) {
149: if (Q-P < 1) return A;
150: Mp = idiv(P+Q,2); M = A[Mp]; B = P; E = Q;
151: while (1) {
152: while (A[B] < M) B++;
153: while (A[E] > M && B <= E) E--;
154: if (B >= E) break;
155: else { T = A[B]; A[B] = A[E]; A[E] = T; E--; }
156: }
157: if (E < P) E = P;
158: if (Level < LevelMax) {
159: /* launch new servers if necessary */
160: if (Proc1 == -1) Proc1 = ox_launch(0);
161: if (Proc2 == -1) Proc2 = ox_launch(0);
162: /* send the requests to the servers */
163: ox_rpc(Proc1,"quickSort",A,P,E,Level+1);
164: ox_rpc(Proc2,"quickSort",A,E+1,Q,Level+1);
165: if (E-P < Q-E) {
166: A1 = ox_pop_local(Proc1);
167: A2 = ox_pop_local(Proc2);
168: }else{
169: A2 = ox_pop_local(Proc2);
170: A1 = ox_pop_local(Proc1);
171: }
172: for (I=P; I<=E; I++) A[I] = A1[I];
173: for (I=E+1; I<=Q; I++) A[I] = A2[I];
174: return(A);
175: }else{
176: /* everything is done on this server */
177: quickSort(A,P,E,Level+1);
178: quickSort(A,E+1,Q,Level+1);
179: return(A);
180: }
181: }
182: \end{verbatim}
183:
184: Another example is a parallelization of the Cantor-Zassenhaus
185: algorithm for polynomial factorization over finite fields. Its
186: fundamental structure is similar to that of quicksort. By choosing a
187: random polynomial, a polynomial is divided into two sub-factors with
188: some probability. Then each subfactor is factorized recursively. In
189: the following program, one of the two sub-factors generated on a server
190: is sent to another server and the other subfactor is factorized on the server
191: itself.
192: \begin{verbatim}
193: /* factorization of F */
194: /* E = degree of irreducible factors in F */
195: def c_z(F,E,Level)
196: {
197: V = var(F); N = deg(F,V);
198: if ( N == E ) return [F];
199: M = field_order_ff(); K = idiv(N,E); L = [F];
200: while ( 1 ) {
201: W = monic_randpoly_ff(2*E,V);
202: T = generic_pwrmod_ff(W,F,idiv(M^E-1,2));
203: if ( !(W = T-1) ) continue;
204: G = ugcd(F,W);
205: if ( deg(G,V) && deg(G,V) < N ) {
206: if ( Level >= LevelMax ) {
207: /* everything is done on this server */
208: L1 = c_z(G,E,Level+1);
209: L2 = c_z(sdiv(F,G),E,Level+1);
210: } else {
211: /* launch a server if necessary */
212: if ( Proc1 < 0 ) Proc1 = ox_launch();
213: /* send a request with Level = Level+1 */
1.3 ! noro 214: /* ox_c_z is a wrapper of c_z on the server */
1.2 noro 215: ox_cmo_rpc(Proc1,"ox_c_z",lmptop(G),E,
216: setmod_ff(),Level+1);
217: /* the rest is done on this server */
218: L2 = c_z(sdiv(F,G),E,Level+1);
219: L1 = map(simp_ff,ox_pop_cmo(Proc1));
220: }
221: return append(L1,L2);
222: }
223: }
224: }
225: \end{verbatim}
226:
227:
228:
229:
230:
231:
232:
FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>