Skip to Main Content
More and more data centers are built using commodity TCP/IP and Ethernet networks because of their low cost and ease-of-use. But TCP would experience catastrophic throughput collapse in a communication pattern, termed "Incast", where multiple senders simultaneously send data to a receiver. This phenomenon potentially arises in many typical applications, such as distributed storage, Map Reduce and web search. However, solutions proposed by former research either are not easy to implement or high in cost. In this paper, we present a simple and cost-effective method - shrinking MTU - to mitigate TCP In cast throughput collapse. Through analysis and extensive simulations, decreasing MTU size can delay the onset of In cast.