AAIC: Adaptive-Sliding-Connection-Window Solution to TCP Incast from Application Layer

CCMA: A Dynamical Concurrent-Connection Management Agent to Mitigate TCP Incast in Datacenters

“…However, these approaches all assume that available bandwidth is known by receiver, which is usually invalid in real-world datacenters. To overcome this drawback, we propose AAIC [25], which adapts to unknown and changing available bandwidth via dynamically adjusting awnd sizes and concurrency number based on data transfer status. Nevertheless, since AAIC (as well as [26]- [29]) is built into each aggregator application, it is inherently unable to control Incast congestion caused by multiple coexisting aggregators on the same receiver.…”

“…In extreme cases, we may completely prevent Incast through setting n(t) = 1 and x(i, t) = (B + C • D) − Y (t) (that is, establishing only one concurrent connection and letting its in-flight packet amount equal to the spare bandwidth, as [27] did). Of course, such simple adjustment strategy often leads to poor goodput [25]. In Section IV, we will discuss how to adjust x(i, t) and n(t) dynamically to achieve both low Incast probability and high goodput.…”

“…We consider various aggregation scenarios, including the ones with numerous workers, background TCP, multiple coexisting aggregators, intermediate bottleneck links, and higher link capacities. The results show that CCMA achieves zero Incast probability and 7.8 ∼ 13.5 times higher goodput than TCP (while DCTCP [10] is 0.8 ∼ 1.1 times higher, RCC [15] is 2.1 ∼ 8.2 times higher, and AAIC [25] is 3.0 ∼ 7.9 times higher). CCMA also greatly reduces the average Flow Completion Time (FCT) of TCP by up to 98.8% for delay-sensitive aggregators.…”

“…3) We evaluate CCMA on both real-world testbed and NS3 simulation. Across a wide range of aggregation scenarios (with numerous workers, background TCP traffic, multiple coexisting aggregators, intermediate bottleneck links, high link capacity, and delay-sensitive aggregators), CCMA maintains zero Incast probability, improves TCP's goodput by 7.8 ∼ 13.5 times, and reduces TCP's average flow completion time by up to 98.8%, which greatly outperforms three state-of-art Incast solutions in [10], [15], and [25].…”

“…However, the solutions in [15]- [24] have relatively high deployment risk, for they non-trivially modify end-host communication software and may incur new issues like incompatibility with other communication hardware (e.g., network infrastructure) or software (e.g., transport protocol). Application-layer solutions such as [25]- [30] are built into each aggregator application. They can be easily deployed in the form of application upgrades.…”