Catch the whole lot in an action: rapid precise packet loss notification in data centers

Abstract

An increasing number of TCP performance issues including TCP Incast, TCP Outcast, and long query completion times are common in large-scale data centers. We demonstrate that the root cause of these problems is that existing techniques are unable to maintain self-clocking or to achieve accurate and rapid packet loss notification. We present cutting payload (CP), a mechanism that simply drops a packet's payload at an overloaded switch, and a SACK-like precise ACK (PACK) mechanism to accurately inform senders about lost packets. Experiments demonstrate that CP successfully addresses the root cause of TCP performance issues. Furthermore, CP works well with other TCP variants used in data center networks.