Poison comes in small packages: Application-driven Reexamination of Datacenter Microbursts
Additional information
Authors
Hosseini M.,
Darabi S.,
Pasandi H. B.,
Nakhjiri M.,
Eugster P. T.
Type
Journal Article
Year
2025
Language
English
Abstract
In the early 2010s, the dramatic increase in the bandwidth of datacenter network links led to the emergence of microsecond-scale congestion events, a challenging phenomenon known as microbursts. Although microbursts may seem very short and small, they pose a significant challenge and can have a severe negative impact on the performance of datacenter networks. Despite the various methods proposed to mitigate microbursts, recent studies show that microbursts and their negative effects still persist in datacenters. In this paper, we show that the primary reason for observing low performance from microburst mitigation methods is that these methods are used in situations or scenarios where they are not suitable or effective. We first discuss this issue by dissecting various microburst mitigation approaches, demonstrating that each method's effectiveness is limited to microbursts with particular characteristics, rather than being a comprehensive solution. Then, we provide a comprehensive measurement and analysis of the diverse characteristics of microbursts, with particular attention to various applications' influences. While intermediate and lower layers of the network stack are predominantly identified as the primary origins of microbursts, our findings reveal that microbursts generated in different applications exhibit different characteristics, which can significantly affect the performance of various mitigation methods. This indicates that optimal microburst mitigation in datacenters requires careful deployment of techniques based on the characteristics of hosted applications. As illustrative examples, through our analysis, we identify which microburst mitigation techniques are effective and which are ineffective for several widely-used datacenter applications.
Journal
Proceedings of the ACM on Measurement and Analysis of Computing Systems
Volume
9
Number ( Month )
2
Pages (or article number)
1-23
ISSN
2476-1249
DOI
Diffusion
License
License undefined
Visibility
Private
