| Autor: |
Flich, J., López, P., Sancho, J. C., Robles, A., Duato, J. |
| Zdroj: |
High Performance Computing (9783540436744); 2006, p49-63, 15p |
| Abstrakt: |
InfiniBand is very likely to become the de facto standard for communication between nodes and I/O devices (SANs) as well as for interprocessor communication (NOWs). The InfiniBand Architecture (IBA) defines a switch-based network with point-to-point links whose topology is arbitrarily established by the customer. Often, the interconnection pattern is irregular. Up*/down* is the most popular routing scheme currently used in NOWs with irregular topologies. However, the main drawbacks of up*/down* routing are the unbalanced channel utilization and the difficulties to route most packets through minimal paths, which negatively affects network performance. Using additional virtual lanes can improve up*/down* routing performance by reducing the head-of-line blocking effect, but its use is not aimed to remove its main drawbacks. In this paper, we propose a new methodology that uses a reduced number of virtual lanes in an efficient way to achieve a better traffic balance and a higher number of minimal paths. This methodology is based on routing packets simultaneously through several properly selected up*/down* trees. To guarantee deadlock freedom, each up*/down* tree is built over a different virtual network. Simulation results, show that the proposed methodology increases throughput up to an average factor ranging from 1.18 to 2.18 for 8, 16, and 32-switch networks by using only two virtual lanes. For larger networks with an additional virtual lane, network throughput is tripled, on average. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
