Distributed scheduling algorithms to improve the performance of parallel data transfers

Abstract

The cost of data transfers, and in particular of I/O operations, is a growing problem in parallel computing. This performance bottleneck is especially severe for data-intensive a p p l ications such a s m ultimedia information systems, databases, and Grand Challenge problems. A promising approach to alleviating this bottleneck i s t o s c hedule parallel I/O operations explicitly.Although centralized algorithms for batch s c heduling of parallel I/O operations have previously been developed, they are not be appropriate for all applications and architectures. We develop a class of decentralized algorithms for scheduling parallel I/O operations, where the objective is to reduce the time required to complete a given set of transfers. These algorithms, based on edge-coloring and matching of bipartite graphs, rely upon simple heuristics to obtain shorter schedules. We present s i m ulation results indicating that the best of our algorithms can produce schedules whose length is within 2 -20% of the optimal schedule, a substantial improvement on previous decentralized algorithms. We discuss theoretical and experimental work in progress and possible extensions.

Block-parallel data analysis with DIY2

“…Durand et al [11] proposed I/O scheduling based on graph matching and coloring. Modeling the set of clients (compute nodes) as one side of a bipartite graph and the set of servers (disks) as the other side, the resulting schedule of I/O transfers attempted to compute a near-optimal schedule.…”

Near-Optimal, Distributed Edge Colouring via the Nibble Method

“…In a distributed setting, the edge colouring problem can be used to model certain types of jobshop scheduling, packet routing, and resource allocation problems. For example, the problem of scheduling I/O operations in some parallel architectures can be modeled as follows [12,7]. We are given a bipartite graph G = (P, R, E) where, intuitively, P is a set of processes and R is a set of resources (say, disks).…”

Distributed Scheduling of Parallel I/O in the Presence of Data Replication

“…As a result the AP algorithm under a particular configuration may have a normalized schedule length greater than 1. Figure 4 compares the normalized schedule lengths from the three algorithms, HDLWF, AP and Random under different data-to-IO-server ratios (8,16,32 and 64 respectively). HDLWF is superior to Random in all cases.…”

“…Durand, Jain and Tseytlin also propose randomized, distributed edge coloring algorithms for bipartite graphs [8,9].…”