Removing the Latency Overhead of the ITB Mechanism in COWs with Source Routing

Clusters of workstations (COWs) are becoming increasingly popular as a cost-effective alternative to parallel computers. In previous papers we presented the in-transit buffer mechanism (ITB) to improve network performance, applying it to COWs with irregular topology and source routing. This mechanism considerably improves the performance of this kind of networks when compared to current source routing algorithms, however it introduces a latency penalty. Moreover, an implementation of this mechanism was performed, showing that the latency overhead of the mechanism may be noticeable, especially for short messages and at low network loads. In this paper, we analyze in detail the latency overhead of ITBs proposing several mechanisms in order to reduce, hide, and remove it. Firstly, we show by simulation the effect of an ITB implementation much slower than the one implemented. Then, we propose three mechanisms that will try to overcome the latency penalty. All the mechanisms are simple and can be easily implemented. Also they are out of the critical path of the ITB packet processing procedure. Results show a very good behavior of the proposed mechanisms, reducing considerably, and even removing the latency overhead.