Understanding the Impact of Removing TCP Binary Exponential Backoff in Data Centers

Abstract

TCP Timeouts are the primary impairment that hurts throughput in data centers. Binary Exponential Backoff (BEB) algorithm is invoked to control interval between consecutive timeouts. In this paper we explore the impact of removing BEB algorithm from TCP on throughput. Our analysis and simulation results show that removing BEB algorithm, even in the case of lower , cannot advance the onset of incast collapse; when increasing SRU size it can noticeably benefit from removing BEB.

MIP: Minimizing the idle period of data transmission in data center networks

“…It's demonstrated in [15] that disabling binary exponential backoff does not advance the onset of incast, instead it can even benefit. We disable the RBEB scheme in TCP and obtain near-optimal MCT from the observation of experiment.…”

“…1) Disabling RBEB: Under severe congestion more than 50% of TCP transmission timeouts invoke RBEB in data center [15]. It's demonstrated in [15] that disabling binary exponential backoff does not advance the onset of incast, instead it can even benefit.…”

A survey on TCP Incast in data center networks

“…Removing BEB can mitigate the Incast problem, resulting in smaller dispersion of response time for all servers to return their portion of data to the client [19]. The simulation experiment results in [19] show that it does not advance the onset of Incast collapse. Instead, it can even benefit from it when using larger SRU size and lower RTOmin.…”

“…Reference [19] points out that removing the TCP binary exponential backoff (BEB) can benefit throughout. Because under severe congestion, more than 50% timeouts can invoke BEB, and then make the server stalled; the link is underutilized during the waiting time for the stalled server to recover.…”

“…[1] points out that altering exponential back off behavior has little impact on mitigating the Incast, it did not describe its NS2 simulation environment to provide convincing result. In [19], it works well under severe congestion (more than 256 servers). However, it has little effect in middle and slight congestion.…”