End-host based mechanisms for implementing flow scheduling in GridNetworks

In Grids, data transfers and network resources need to be managed in a more deterministic way than in the Internet. New approaches like flow scheduling are proposed and studied as alternatives to traditional QoS and reservation proposals. To enable such flow scheduling approaches, runtime mechanisms controlling flow sending time and rate have to be implemented in the data plane. This paper quantifies and compares such end-host based mechanisms combined with transport protocols to instantiate different scheduling strategies in a range of latency conditions. We show that, a single-rate scheduling strategy implemented by an AIMD-based protocol and a packet pacing mechanism offers predictable performance and is insensitive to latency. This paper also highlights the limits of other strategies and rate limitation mechanisms like token bucket which generates unpredictability and other drawbacks.