The scale of modern data centers enables developers to deploy applications across thousands of servers. The variety of applications and the scale of operations impose onerous challenge of meeting application performance requirements while maintaining efficiency. Today, data center operators typically over-provision the network and run services at low utilization to rein in latency outliers, thus decreasing efficiency. This large scale inefficiency results in high monetary, energy, and management expenses. This dissertation focuses on redesigning the end-host network stack to improve network efficiency and achieve low latency and latency variation at high utilizations . We begin by studying traffic emanating from modern servers across a variety of data center applications. We find that traffic is highly bursty, which contradicts the network flow model where traffic is uncorrelated. We use this observation to design networks that can benefit from bursty behavior. Second, in data center applications, predictability in service time and controlled latency, especially tail latency, are essential for building performant applications. Current practice has been to run such services at low utilization to rein in latency outliers, which decreases efficiency. To combat this, we present Chronos, which is a framework to reduce end-host latency and latency variation. Chronos reduces Memcached latency by a factor of 20 compared to typical deployments. Third, a range of new data center switch designs incorporating wireless or optical circuits depend on fast reconfiguration of the underlying topology. These hybrid designs assume a perfect, closed-loop control plane which end-host network stacks cannot provide today. We present the design and implementation of a closed-loop control plane using only software changes at the end host operating system that enable these topologies to support unmodified applications running over TCP. Taken together, these contributions demonstrate we can meet performance requirements of data center applications while running data centers at high levels of efficiency

Solving "Big Data" problems requires bridging massive quantities of compute, memory, and storage, which requires significant amounts of bisection bandwidth. Topologies like Fat-tree, VL2, and HyperX achieve a scale-out design by leveraging multiple paths from source to destination. However, traditional routing protocols are not able to effectively utilize these links while also harnessing spare network capacity to better statistically multiplex network resources. In this work we present Dahu, a switch mechanism for efficiently load balancing traffic in multipath networks. Dahu avoids congestion hot-spots by dynamically spreading traffic uniformly across links, and forwarding traffic over non-minimal paths where possible. By performing load balancing primarily using local information, Dahu can act more quickly than centralized approaches, and responds to failure gracefully. Our evaluation shows that Dahu delivers up to 50% higher throughput relative to ECMP in an 8,192 server Fat-tree network and up to 500% improvement in throughput in large scale HyperX networks with over 130,000 servers.

Pre-2018 CSE ID: CS2013-0992

Engineering large-scale data center applications built from thousands of commodity nodes requires both an underlying network that supports a wide variety of traffic demands, and low latency at microsecond timescales. Many ideas for adding innovative functionality to networks, especially active queue management strategies, require either modifying packets or performing alternative queuing to packets in-flight on the data plane. However, configuring packet queuing, marking, and dropping is challenging, since buffering in commercial switches and routers is not programmable. In this work, we present NetBump, a platform for experimenting with, evaluating, and deploying a wide variety of active queue management strategies to network data planes with minimal intrusiveness and at low latency. NetBump leaves existing switches and endhosts unmodified by acting as a ``bump on the wire,'' examining, marking, and forwarding packets at line rate in tens of microseconds to implement a variety of virtual active queuing disciplines and congestion control mechanisms. We describe the design of NetBump, and use it to implement several network functions and congestion control protocols including DCTCP and 802.1Qau quantized congestion notification.

Pre-2018 CSE ID: CS2012-0976

In datacenter applications, predictability in service time and controlled latency, especially tail latency, is essential for building performant applications. This is especially true for applications or services build by accessing data across thousands of servers to generate a user response. Current practice has been to run such services at low utilization to rein in latency outliers, which decreases efficiency and limits the number of service invocations developers can issue while still meeting tight latency budgets. In this paper, we analyze the three datacenter applications, Memcached, OpenFlow, and web search, to measure the effect on tail latency of 1) kernel socket handling, NIC interaction, and the network stack, 2) application locks contested in the kernel, and 3) application-layer queuing due to requests being stalled behind straggler threads. We propose a novel approach of reducing the above sources of latency by relying on support from the NIC hardware, and we find that the resulting improvements indeed dramatically reduce end-to-end application latency.

Pre-2018 CSE ID: CS2012-0977

"Big Data" computing increasingly utilizes the MapReduce programming model for scalable processing of large data collections. Many MapReduce jobs are I/O-bound, and so minimizing the number of I/O operations is critical to improving their performance. In this work, we present ThemisMR, a MapReduce implementation that reads and writes data records to disk exactly twice, which is the minimum amount possible for data sets that cannot fit in memory. In order to minimize I/O, ThemisMR makes fundamentally different design decisions from previous MapReduce implementations. ThemisMR performs a wide variety of MapReduce jobs – including click log analysis, DNA read sequence alignment, and PageRank – at nearly the speed of TritonSort’s record-setting sort performance.

Pre-2018 CSE ID: CS2012-0983