We introduce FlowComb, a network management framework
that helps Big Data processing applications, such as
Hadoop, achieve high utilization and low data processing
times. FlowComb predicts application network transfers,
sometimes before they start, by using software agents installed
on application servers and while remaining completely
transparent to the application. A centralized decision
engine collects data movement information from
agents and schedules upcoming flows on paths such that
the network does not become congested. Results on our
lab testbed show that FlowComb is able to reduce the
time to sort 10GB of randomly generated data by 35%
while changing paths for only 6% of the transfers.