Telcordia Technologies

About: Telcordia Technologies is a based out in . It is known for research contribution in the topics: Network packet & Node (networking). The organization has 3097 authors who have published 4737 publications receiving 237882 citations.

Drop from front of buffer policy in feedback networks

Abstract: A method for transmitting packets or cells (or both) in a communications network (10) is disclosed. The packets or cells are transmitted along a forward communications path in a network from a source node (h1) via one or more intermediary nodes (r1, x1, r5 and r2) to a destination node (h4). At the intermediary nodes (r1, x1, r5 or r2), the packets or cells are received in a buffer (38 or 42-1 to 42-4). The packets or cells are transmitted along the forward communications path according to a communications schedule. In the presence of congestion at one of the intermediary nodes (x1), an indication of the congestion (302-342) is provided to the destination nodes (h4) of the first packets to be transmitted according to the schedule. An indication of the congestion of the first packets or cells is provided by the destination nodes (h4) of the first packets or cells to the source nodes (h1) of the first packets or cells via a feedback communications path.

Optimal disk I/O with parallel block transfer

Abstract: 1 I n t r o d u c t i o n Sorting is the canonical information-processing application. It accounts for roughly 20-25 percent of the computing resources on large-scale computers [8,9]. In most applications the file of records cannot fit into internal memory and must be stored on (external) sec* S u p p o r t was p rov ided in pa r t by an NSF Pres iden t i a l Young Inves t iga to r Award C C R 8 9 0 6 4 1 9 wi th m a t c h i n g f u n d s f rom IBM, by NSF resea rch g r a n t D C R 8 4 0 3 6 1 3 , a n d by O N R g ran t N 0 0 0 1 4 8 3 K 0 1 4 6 , A R P A Orde r No. 4786. t T h i s work was done in pa r t while t he a u t h o r was a t Brown Univers i ty a n d s u p p o r t e d by a Bellcore g r a d u a t e fellowship. Permission to copy without fee all or part of this matertial is granted provided that the copies are not made or distributed for direct commercial advantage, the ACM copyright notice and the title of the publication and its date appear, and notice is given that copying is by permission of the Association for Computing Machinery. To copy otherwise, or to republish, requires a fee and/or specific permission. ondary storage, usually in the form of magnetic or optical disks. Sorting in this framework is called ezternal sorting. The bottleneck in external sorting and many other applications is the time for the input /output (I/O) between internal memory and the disks. This bottleneck is accentuated as processors get faster and parallel computers are used. One remedy is to use secondary storage systems with parallel capabilities. Previous work on I /O efficiency has concentrated on two-level and multilevel models. (Further references can be found in the papers mentioned below.) Aggarwal and Vitter [3] present optimal upper and lower bounds for sorting-related problems using a two-level memory model in which P blocks, each consisting of B contiguous records, can be transferred simultaneously in a single I/O. Their model is somewhat unrealistic, however, because secondary storage is usually partitioned into separate physical devices, each capable of transferring only one block per I/O. Multilevel hierarchical memory models are introduced in [1,2], one of them taking into account block transfer. Access to a location z takes time f(z), and in the blocking version, access to successive locations takes one unit of time per location. Optimal bounds are obtained for several problems. Parallel transfer using several hierarchies simultaneously was not considered in [1,2]. In this paper we are interested in optimal algorithms for powerful two-level and hierarchical memory models that allow P simultaneous transfers of data. To be realistic, we require that each block transfer must access a separate secondary storage device. In Section 3 we define a realistic two-level memory model with parallel block transfer, and we state our main results, which give tight upper and lower bounds on the number of I /Os needed to solve several important problems. The problems, which include sorting, permuting, matrix transposition, FFT, permutation networks, and standard matrix multiplication, are defined in Section 2. The two-level model corre© 1990 ACM 089791-361-2/90/0005/0159 $1.50 159 sponds to having an internal memory and P disks, each capable of simultaneously transferring one block of B records. Our measure of performance is the number of parallel I /Os required; this ignores internal computation time, but the internal processing done by our algorithms is simple enough so that in practice it can be overlapped with the I /O time. Our algorithms can be significantly faster than those obtained by the wellknown technique of disk striping. The restriction that only one block can be accessed per disk during an I /O is what distinguishes our model from the less realistic model of [3]. This distinction is akin to the difference in parallel computation between the more realistic MPC (module parallel computer) model and the less realistic PRAM model. The algorithms we develop on our more realistic model use the same number of I /Os as those in [3] for the less realistic model. In Section 4 we define two uniform memory models, each consisting of P hierarchical memories connected together at their base levels. The P hierarchical memories are of the type discussed in [1,2]. We give optimal time bounds for the problems in each model. Sections 5-7 are devoted to the algorithms and analysis for the two-level model. In Section 5 we develop optimal algorithms for matrix transposition, FFT, and permutation networks, by making use of the shuffle-merge primitive. Even though these problems are sorting-related, it is much easier to develop optimal algorithms for them than it is for sorting, since their I /O schedules are nonadaptive. Our main result is the optimal randomized algorithm for sorting (and permuting) and its probabilistic analysis in Section 6. The probability that it uses more than ~ times the optimal number of I /Os is exponentially small in ~(logt)log(M/B), where i is the internal memory size. The sorting algorithm is a variant of distribution sort; a combination of two interesting randomized techniques is used to do the partitioning so as to take full advantage of parallel block transfer. In Section 7 we cover standard matrix multiplication. In Section 8 we show how to apply the algorithms developed for the two-level model to get optimal algorithms for the hierarchical models. The hierarchical algorithms are optimal because the internal processing in the corresponding two-level algorithms is efficient. For simplicity of notation, we use logx, where x >__ 1, to denote the quantity max{l , log 2 x}. We also assume for the most part that the numbers we deal with are integers and divide evenly, so as not to clutter up the manuscript with numerous floors and ceilings. On the other hand, there are some places in the analysis where this issue is important , so we will be more precise in those cases. The complete set of results and the full proofs appear in the full version of this paper [15]. 2 P r o b l e m Def ini t ions The sorting problem consists of rearranging the N records so that their key values are in nondecreasing order. The permuting problem is the special case of sorting in which the key values of the N records are distinct integers in {1, 2, . . . , N}. Matrix transposition is the special case of permuting in which the permutation to be realized corresponds to matr ix transposition. The FFT problem can be phrased as the problem of pumping the records into and out of internal memory so as to permit the computation implied by the FFT digraph. The permutation network problem is like the FFT problem, except that a comparator network capable of realizing any permutation is used instead of the FFT digraph. There is an important difference between permutation networks and general permuting. In the latter case, the particular I /Os performed may depend upon the desired permutation, whereas with permutation networks all N! permutations can be generated by the same sequence of I/Os. The standard matrix multiplication problem deals with computing the product C = A x B of two k x k matrices, by following the well-known algorithm that uses O(k 3) additions and multiplications. The elements of A, B, and C are assumed to be in row-major order. 3 Two-Leve l M e m o r y M o d e l First we define the parameters for our two-level memory model (or disk model) with parallel block transfer: Def in i t i on 1 The parameters are defined by N = ~ records in the file; M = # records that can fit in internal memory; B = ~ records per block;

Serving network selection and multihoming using ip access network

Abstract: In some illustrative embodiments, an IP-layer based network selection and multihoming method is provided that enables a flexible and secure dynamic selection of one or more serving networks for use by a client node. The method is independent of any link-layer technology. A serving network can be an ISP network, a NAP network exchange facility, a VLAN, or the like. Network information is advertised to a client node, the client node is authenticated and authorized for use of an access router, and a secure tunnel is established between the client node and the access router. The method can be implemented by using standard protocols, and can work over any existing or future link-layer technologies that are able to carry IP datagrams, without any modification.

Varying approaches to topical web query classification

Abstract: Topical classification of web queries has drawn recent interest because of the promise it offers in improving retrieval effectiveness and efficiency. However, much of this promise depends on whether classification is performed before or after the query is used to retrieve documents. We examine two previously unaddressed issues in query classification: pre versus post-retrieval classification effectiveness and the effect of training explicitly from classified queries versus bridging a classifier trained using a document taxonomy. Bridging classifiers map the categories of a document taxonomy onto those of a query classification problem to provide sufficient training data. We find that training classifiers explicitly from manually classified queries outperforms the bridged classifier by 48% in F1 score. Also, a pre-retrieval classifier using only the query terms performs merely 11% worse than the bridged classifier which requires snippets from retrieved documents.

Optical fiber amplifier with flattened gain

Abstract: An apparatus and method for flattening the gain of an optical fiber amplifier (18) doped with an ion such as erbium and which is pumped by a pump laser (20). Optical couplers (32, 34) inserted before and after the fiber amplifier couple the main optical path to an optical ring passing through the fiber amplifier so as to form a ring laser. An optical isolator (32) placed in the ring causes the lasing light to only counterpropagate relative to the optical signal being amplified. When the fiber amplifier primarily exhibits inhomogeneous broadening and the ring lases, the lasing light clamps the gain to a value determined by the loop loss, and the value of the clamped gain is relatively uniform across a wide bandwidth.