## Efficient dispersal of information for security, load balancing, and fault tolerance (1989)

Venue: | Journal of the ACM |

Citations: | 484 - 1 self |

### BibTeX

@ARTICLE{Rabin89efficientdispersal,

author = {Michael Rabin},

title = {Efficient dispersal of information for security, load balancing, and fault tolerance},

journal = {Journal of the ACM},

year = {1989},

volume = {36},

pages = {335--348}

}

### Years of Citing Articles

### OpenURL

### Abstract

Abstract. An Information Dispersal Algorithm (IDA) is developed that breaks a file F of length L = ( F ( into n pieces F,, 1 5 i 5 n, each of length ( F, 1 = L/m, so that every m pieces suffice for reconstructing F. Dispersal and reconstruction are computationally efficient. The sum of the lengths ( F, 1 is (n/m). L. Since n/m can be chosen to be close to I, the IDA is space eflicient. IDA has numerous applications to secure and reliable storage of information in computer networks and even on single disks, to fault-tolerant and efficient transmission of information in networks, and to communi-cations between processors in parallel computers. For the latter problem provably time-efftcient and highly fault-tolerant routing on the n-cube is achieved, using just constant size buffers. Categories and Subject Descriptors: E.4 [Coding and Information Theory]: nonsecret encoding schemes

### Citations

1897 | How to share a secret
- Shamir
- 1979
(Show Context)
Citation Context ...nt 1, Pr(more than 0.25n paths D,(x) fail) 5 The same bound holds for the paths E,(x). By our analysis, the probability that Px will not be reconstructable at r(x) is at most twice the probability in =-=(11)-=-. Since there are N = 2” packets, the probability that some packet will fail within a transmissjon round is at most 2 ’ N . (4e/n)0.25’r. Although asymptotically the above probability of failure is sm... |

498 |
The connection machine
- Hillis
- 1989
(Show Context)
Citation Context ... # 0; iZj+xi#xj and yifyj, (this requires n + m < p). Define ( 1 1 ai=- - 15isr.z. Xi + y, ’ ’ ’ ’ ’ X; + ym 1 ’ Let A be the matrix with rows aI, . . . , a,,, then [5, p. 351 It follows from (3) and =-=(4)-=- that 1 A 1 # 0. Thus any m vectors in (a,, . . . , a,) are linearly independent. Furthermore, if A-* = [bij], then bij = (-l)i+‘(.4(j, i)(/]A 1, where A(j, i) is the matrix obtained from A by deletin... |

457 |
Algebraic Coding Theory
- Berlekamp
- 1968
(Show Context)
Citation Context ...is no better than n-fold replication. On the other hand, in Shamir’s scheme, any m - 1 pieces give no information about 9. Our method can be viewed as belonging to the field of error correction codes =-=[2]-=-, in which extra bits are added to a message creating a block, so that after occurrence of any k errors within the block, the message can still be reconstructed. These k errors can be anywhere in the ... |

282 |
Probabilistic construction of deterministic algorithms: approximating packing integer programs
- Raghavan
- 1988
(Show Context)
Citation Context ...,x,t)=n. (7) hE c n From the definition of the RA it follows that the random variables Y( y, x, t ) are pairwise independent (here y and t are fixed, x E C,). A recent theorem of Raghavan and Spencer =-=[9]-=- states that if Y,, . . . , Yhi are independent Bernoulli trials such that the expected value E for their sum is andif6>O,then(withe=2.71 . ..). E CY; =n, (i ) The above analysis shows that the condit... |

245 |
Fingerprinting by random polynomials
- Rabin
- 1981
(Show Context)
Citation Context ...eplace a piece Fi at some node by a string G. If we use G as one of the pieces for reconstruction, we shall not get F back. A reliable method to counter this replacement attack is to use fingerprints =-=[8]-=-. We randomly choose an irreducible polynomial f(x) E Z,,[x] of a small degree k, say, k = 11. We consider the piece (including the header ai) F; = djl . . . d[M, A4 = N/m + m, as a polynomial F;(X) =... |

216 |
The cosmic cube
- Seitz
- 1985
(Show Context)
Citation Context ...ity of overflow of some buffer at some time 1 I t 5 2(~ + l), that is, during a transmission round is at most Pr(some overflow during one transmission round) 5 24” . 1.p .N.2(n+l)=N4. 2(n + 1) 18,1 . =-=(10)-=- This implies the theorem for n 2 4. By way of example, for n = 10, the probability, for a transmission round, of even one packet Px not arriving is, according to (lo), at most 6 . 10hi4. We now turn ... |

215 | A scheme for fast parallel communication - Valiant - 1982 |

178 |
Combinatorial Theory
- Hall
- 1967
(Show Context)
Citation Context ... simplicity we assume that the dimension n of the cube is itself a power of 2 or that n - 1 is a prime, n = 0 mod 4 (e.g., n = 12, 24). For such n, and many others there exist n x n Hadamand matrices =-=[3]-=- H,, that are very rapidly calculable. Condition (12) implies that two different rows in H, differ in exactly n/2 locations.Delinev,EC,,, 1 sisnbyv,[jJ= lifd,,= l,v,[j]=Oifd,,=-1. Then d(v,, v,) = n/2... |

84 | Probabilistic algorithms in finite fields
- Rabin
- 1980
(Show Context)
Citation Context ...es we can directly use the field E = GF(28) of characteristic 2 and having 256 elements. All we need is an irreducible polynomial p(x) E Z2[x] of degree 8 to allow us to effectively compute in E (see =-=[7]-=-). 2. The Independence Condition Let xl, . . . , x,, yI, . . . , y,,, E Z, satisfy the conditions: For all i and j xi + yj # 0; iZj+xi#xj and yifyj, (this requires n + m < p). Define ( 1 1 ai=- - 15is... |

66 |
An Introduction to Linear Algebra
- Mirsky
- 1990
(Show Context)
Citation Context ...nsmit the pairs (E(P,), Fi), I 5 i 5 n. If we use IDA for transmission we must also transmit E(f). When receiving pairs, call them (H, G), for reconstructing F, we test whether E-‘(H) = res(G(x), f). =-=(5)-=- Only if (5) holds for a pair (H, G), do we use G as one of the pieces Fi. By assumption, we shall have m such pieces. It is proved in [8] that an adversary, who does not know the polynomial f(x), has... |

22 |
Parallel communication with limited buffers
- Pippenger
- 1984
(Show Context)
Citation Context ...e units is shown to be smaller than N-(2K-‘). The local queues, in which packets await their transfer to neighboring nodes, will not contain more than K . logzN packets, with the same probability. In =-=[6]-=-, Pippenger showed that fixed-size queues sufftce, at the expense of lengthening the transmission time. We employ the IDA to achieve assured fast transmission time, small fixed (i.e., independent of N... |

12 |
A model of SIMD machines and a comparison of various interconnection networks
- Siegel
- 1979
(Show Context)
Citation Context ...ted to fixed locations that are mutually far apart, and then dispatched to random intermediate destinations. A matrix H = [d,] is Hadamand if d,, = fl and i # j implies d;,d,, + . . f + d,,,d,,, = 0. =-=(12)-=- For the sake of simplicity we assume that the dimension n of the cube is itself a power of 2 or that n - 1 is a prime, n = 0 mod 4 (e.g., n = 12, 24). For such n, and many others there exist n x n Ha... |

6 |
Pooling splitting and restituting information to overcome total failure of some channels of communication
- Asmuth, Blakely
- 1982
(Show Context)
Citation Context .... Thus F= (b,, . . . . b,), (bm+,, . . . . b2m), . . . . Denote S, = (b,, . . . , b,), etc. For i = 1, . . . , n, where Fi = Gil, Ci2, * -. 7 ciN/m, Cik = Ui . Sk = Uil * b<k-I)m+l + .*. + Ui, * bkm. =-=(1)-=- It follows that 1 Fi 1 = 1 F 1 /m. If m pieces of F, say, F, , . . . , F, are given, we reconstruct F as follows. Let A = (Uii)Il;,j=m be the ~lt X m matrix whose ith row is Ui. It is readily seen th... |