ECHO: Recreating network traffic maps for datacenters with tens of thousands of servers

Abstract

Abstract-Large-scale datacenters now host a large part of the world's data and computation, which makes their design a crucial architectural challenge. Datacenter (DC) applications, unlike traditional workloads, are dominated by user patterns that only emerge in the large-scale. This creates the need for concise, accurate and scalable analytical models that capture both their temporal and spatial features and can be used to create representative activity patterns. Unfortunately, previous work lacks the ability to track the complex patterns that are present in these applications, or scales poorly with the size of the system. In this work, we focus on the network aspect of datacenter workloads. We present ECHO, a scalable and accurate modeling scheme that uses hierarchical Markov Chains to capture the network activity of large-scale applications in time and space. ECHO can also use these models to re-create representative network traffic patterns. We validate the model against real DC-scale applications, such as Websearch and show marginal deviations between original and generated workloads. We verify that ECHO captures all the critical features of DC workloads, such as the locality of communication and burstiness and evaluate the granularity necessary for this. Finally we perform a detailed characterization of the network traffic for workloads in DCs of tens of thousands of servers over significant time frames.

Inside the Social Network's (Datacenter) Network

“…Lacking concrete data, researchers often design for the worst case, namely an all-to-all traffic matrix in which each host communicates with every other host with equal frequency and intensity [4]. Such an assumption leads to the goal of delivering maximum bisection bandwidth [4,23,36], which may be overkill when demand exhibits significant locality [17].…”

Channel reservation protocol for over-subscribed channels and destinations

“…The background hotspot traffic can also represent the collection of intermediate web search results using parts of the search index, performed by numerous leaf nodes and sent to few front end web server nodes, or reduction operations with constructs such as MPI [33]. The benign foreground intra-cluster UR pattern represents multi-purpose traffic that is confined within the same or consecutive clusters or racks [13]. It is this traffic that should remain unaffected by the ill-behaved hotspot traffic.…”

“…In this traffic pattern we collect statistics for benign foreground intra-cluster UR traffic where every source (including the ones participating in the hotspot pattern) transmits at a variable rate only to destinations in the same cluster. The hotspot traffic represents the small number of nodes that communicate with most of the machines in the system (and across clusters), because they are running the task scheduler, aggregators or monitoring systems [13]. The background hotspot traffic can also represent the collection of intermediate web search results using parts of the search index, performed by numerous leaf nodes and sent to few front end web server nodes, or reduction operations with constructs such as MPI [33].…”

“…We simulate realistic traffic observed in HPC systems and datacenters [13,33] with our combined traffic pattern which uses background hotspot traffic where half of the sources in each cluster transmit at full rate to a hotspot in each network cluster (each source transmits to all hotspots). In this traffic pattern we collect statistics for benign foreground intra-cluster UR traffic where every source (including the ones participating in the hotspot pattern) transmits at a variable rate only to destinations in the same cluster.…”

iBench: Quantifying interference for datacenter applications

“…Given this challenge, there is extensive work on characterization and modeling of DC applications [5], [14], [15], [16], [25], [38] that leads to generated workloads with characteristics that closely resemble those of the original application. The generated workloads can then be used in system studies without the limitation of needing access to real DC workloads.…”