SaHaD: A fine-grained in-network latency measurement for datacenter network

Datacenters have become a fundamental infrastructure for modern cloud services. The traffic pattern of DataCenter Network (DCN) however has dramatic difference compared to conventional Internet traffic. In particular, momentary bursts have been widely observed which significantly impacts the utilization, packet latency and packet loss of the links. We have also found from real traces that the datacenter traffic has much shorter packet arrival interval and more short-lived bursts. The packet loss and reordering within one router or through multi-path routing of modern DCN further aggravate the problem. In this paper, we propose SaHaD (Sampling, Hash and Digest), a passive fine-grained in-network latency measurement mechanism for DCN. We define a multiple slot structure and leverage hash function to reduce the influence of packet loss and reordering. We present an adaptive packet sampling algorithm for latency measurement. A segmented digest is further developed to derive packet loss and reordering level with minimized overhead. Our trace-driven experiment suggests that SaHaD achieves high efficiency and accuracy for highly bursty and short-lived DCN traffic, and yet with acceptable deployment and maintenance cost. © 2013 Binary Information Press.