Mugurel Ionut Andreica

Efficient Data Structures for Online QoS-Constrained Data Transfer Scheduling

Distributed applications and services requiring the transfer of large amounts of data have been developed and deployed all around the world. The best effort behavior of the Internet cannot offer to these applications the necessary quality of service (QoS) guarantees, making the development of data transfer scheduling techniques a necessity. In this paper we propose novel methods of efficiently using some well-known data structures (e.g. the segment tree and the block partition), which can be implemented in a resource manager (e.g. grid job scheduler, bandwidth broker) in order to serve quickly large numbers of advance resource reservation and allocation requests.

Maximum reliability k-hop multicast strategy in tree networks

In this paper we consider directed tree networks, for which the reliability of each edge (a real number between 0 and 1) is known. For these networks, we investigate the problem of finding a k-hop multicast strategy of maximum reliability. This problem is equivalent to the k-station placement problem, which was previously solved in polynomial time for trees. We present here O(kldrn2) and O(kldrn3) dynamic programming algorithms for the problem, which improve upon the previous best known solution, which is O(kldrn2ldrlog(n)) and rather complicated to implement. We then extend the algorithms to general directed graphs and also present some new algorithms for this case.

Optimal Offline TCP Sender Buffer Management Strategy

The transmission control protocol (TCP) uses a sliding window in order to enforce flow control. The receiver advertises its available buffer space to the sender, which cannot transmit more data than the advertised space. Transmitted data is first copied from application buffers into TCP buffers and from there it is sent through the network. In this paper we propose a model which characterizes the senderpsilas behavior throughout the duration of a TCP conversation. The model is suitable in the case of powerful, but variably loaded senders, slow receivers, fast connections and moderate amounts of data transmitted. For this model we present an O(nldrlog2n) algorithm which computes the minimum processing time spent by the sender, if the window sizes advertised by the receiver and the senderpsilas load are known in advance. The solution is based on an algorithmic framework for the segment tree data structure, which we introduce in this paper.

Performance evaluation of a Python implementation of the new LEDBAT congestion control algorithm

LEDBAT is a new congestion control algorithm which was proposed in an IETF draft in March 2009. LEDBATs goal is to provide a less-than-best-effort data transport service, i.e. a LEDBAT flow should back-off in the presence of other competing flows (TCP flows in particular). Because of its intended friendliness to competing flows, LEDBAT has already been implemented in the popular Bittorrent client uTorrent. In this paper we present performance evaluation results for our Python implementation of LEDBAT. We ran tests both in emulated networks as well as in real world networks. One of the tests involved approximately 300 volunteers spread throughout the world.

A fault-tolerant peer-to-peer object storage architecture with multidimensional range search capabilities and adaptive topology

In this paper we present a fault-tolerant, collaborative peer-to-peer object storage architecture with adaptive topology and efficient multidimensional range search capabilities. Every stored object has a fixed set of index properties, whose ranges of values form a multidimensional geometric property space. The architecture efficiently supports multidimensional range queries by mapping the peer identifiers into the property space of the stored objects. The potential disadvantage of not balancing the load on the peers which may be caused by this approach in the case of a static topology is addressed by introducing a dynamic topology, which attempts to balance the storage load. The architecture seamlessly supports dynamic node arrivals and departures.

High multiplicity scheduling of file transfers with divisible sizes on multiple classes of paths

Distributed applications and services requiring the transfer of large amounts of data have been developed and deployed world wide. The best effort model of the Internet cannot provide these applications with the so much needed quality of service guarantees, making necessary the development of file transfer scheduling techniques, which optimize the usage of network resources. In this paper we consider the high multiplicity scheduling of file transfers over multiple classes of paths with the objective of minimizing the makespan, when the files have divisible sizes. We also consider another objective, that of maximizing the total profit, in the context of some special types of mutual exclusion constraints (tree and clique constraint graphs).

Reliability analysis of tree networks applied to balanced content replication

The fault tolerance of the communication topology of a distributed system is a very important feature which needs to be analyzed carefully. In this paper we propose a new reliability metric for tree topologies, based on the unrestricted vertex multicut problem on trees, for which we present the first optimal linear time algorithm. We present evaluation results of the reliability metric on tree networks used for balanced content replication. For this problem, we also developed a new O(nmiddotk) algorithm solving the k-equitable coloring problem on trees, based on a hierarchy of color relabeling permutations.

Central placement of storage servers in tree-like content delivery networks

In this paper we consider several problems regarding the optimal placement of storage servers in content delivery networks with tree-like topologies (paths, trees and cacti). We present several novel algorithmic solutions for locating (k+p)-centers and connected k-centers (and k-medians) in these networks, both in the continuous and the discrete cases. We comment both on the theoretical and the practical efficiency of the proposed algorithms.

Towards a Real-Time Scheduling Framework for Data Transfers in Tree Networks

As the amount of transferred data increases world-wide, network data transfers may exhibit poor performance due to resource usage conflicts, even when enough resources are available (at a different place or at another time). Thus, efficient resource management and intelligent scheduling of the data transfers are required. These problems are, however, quite difficult to solve, both from a theoretical and a practical perspective. In this paper we propose several novel algorithmic techniques for the real-time scheduling of data transfers when the network has a tree topology. Some of these techniques can be later implemented in a (re)configurable data transfer scheduling framework.

Replication mechanisms for a distributed time series storage and retrieval service

In this paper we present the prototype architecture of a distributed service which stores and retrieves time series data, together with replication mechanisms employed in order to provide both reliability and load balancing. The entries of each time series are stored locally on the machines running the instances of the service. Each entry is eventually fully replicated on every service instance. Our replication mechanisms depend on whether there is only one service instance receiving each entry of a time series from a client or there may be multiple such instances.