max planck institut
informatik

# MPI-INF or MPI-SWS or Local Campus Event Calendar

 << Previous Entry Next Entry >> New Event Entry Edit this Entry Login to DB (to update, delete)
Title: Multiple Choice Allocations with Small Maximum Loads Megha Khosla Max-Planck-Institut für Informatik - D1 Promotionskolloquium D1, D2, D3, D4, D5, RG1, SWS, MMCIWe use this to send out email in the morning. AG Audience English
Date: Tuesday, 4 March 2014 11:00 30 Minutes Saarbrücken E1 4 024
 The idea of using multiple choices to improve allocation schemes is now well understood and is often illustrated by the following example. Suppose $n$ balls are allocated to $n$ bins with each ball choosing a bin independently and uniformly at random. The \emph{maximum load}, or the number of balls in the most loaded bin, will then be approximately $\log n \over \log \log n$ with high probability. Suppose now the balls are allocated sequentially by placing a ball in the least loaded bin among the $k\ge 2$ bins chosen independently and uniformly at random. Azar, Broder, Karlin, and Upfal showed that in this scenario, the maximum load drops to ${\log \log n \over \log k} +\Theta(1)$, with high probability, which is an exponential improvement over the previous case. In this thesis we investigate multiple choice allocations from a slightly different perspective. Instead of minimizing the maximum load, we fix the bin capacities and focus on maximizing the number of balls that can be allocated without overloading any bin. In the process that we consider we have $m=\lfloor cn \rfloor$ balls and $n$ bins. Each ball chooses $k$ bins independently and uniformly at random. \emph{Is it possible to assign each ball to one of its choices such that the no bin receives more than $\ell$ balls?} For all $k\ge 3$ and $\ell\ge 2$ we give a critical value, $c_{k,\ell}^*$, such that when $cc_{k,\ell}^*$ this is not the case. In case such an allocation exists, \emph{how quickly can we find it?} Previous work on total allocation time for case $k\ge 3$ and $\ell=1$ has analyzed a \emph{breadth first strategy} which is shown to be linear only in expectation. We give a simple and efficient algorithm which we also call \emph{local search allocation} (LSA) to find an allocation for all $k\ge 3$ and $\ell=1$. Provided the number of balls are below (but arbitrarily close to) the theoretical achievable load threshold, we give a \emph{linear} bound for the total allocation time that holds with high probability. We demonstrate, through simulations, an order of magnitude improvement for total and maximum allocation times when compared to the state of the art method. Our results find applications in many areas including hashing, load balancing, data management, orientability of random hypergraphs and maximum matchings in a special class of bipartite graphs.
Name(s): Megha Khosla