The basic building block of ever larger data centers has shifted from a rack to modular container with hundreds or even thousands servers. Delivering scalable bandwidth among such containers is challenge. A number recent efforts promise full bisection between all servers, though significant cost, complexity, and power consumption. We present Helios, hybrid electrical/optical switch architecture that can deliver reductions in the switching elements, cabling, consumption relative recently...

Large cloud service providers have invested in increasingly larger datacenters to house the computing infrastructure required support their services. Accordingly, researchers and industry practitioners alike focused a great deal of effort designing network fabrics efficiently interconnect manage traffic within these performant yet efficient fashions. Unfortunately, datacenter operators are generally reticent share actual requirements applications, making it challenging evaluate practicality...

The basic building block of ever larger data centers has shifted from a rack to modular container with hundreds or even thousands servers. Delivering scalable bandwidth among such containers is challenge. A number recent efforts promise full bisection between all servers, though significant cost, complexity, and power consumption. We present Helios, hybrid electrical/optical switch architecture that can deliver reductions in the switching elements, cabling, consumption relative recently...

Recent proposals have employed optical circuit switching (OCS) to reduce the cost of data center networks. However, relatively slow times (10--100 ms) assumed by these approaches, and accompanying latencies their control planes, has limited its use only largest networks with highly aggregated constrained workloads. As faster switch technologies become available, designing a plane capable supporting them becomes key challenge.

Large cloud service providers have invested in increasingly larger datacenters to house the computing infrastructure required support their services. Accordingly, researchers and industry practitioners alike focused a great deal of effort designing network fabrics efficiently interconnect manage traffic within these performant yet efficient fashions. Unfortunately, datacenter operators are generally reticent share actual requirements applications, making it challenging evaluate practicality...

The ever-increasing bandwidth requirements of modern datacenters have led researchers to propose networks based upon optical circuit switches, but these proposals face significant deployment challenges. In particular, previous dynamically configure switches in response changes workload, requiring network-wide demand estimation, centralized assignment, and tight time synchronization between various network elements--- resulting a complex unwieldy control plane. Moreover, limitations the...

In data center applications, predictability in service time and controlled latency, especially tail are essential for building performant applications. This is true applications or services built by accessing across thousands of servers to generate a user response. Current practice has been run such at low utilization rein latency outliers, which decreases efficiency limits the number invocations developers can issue while still meeting tight budgets.

Sprocket is a highly configurable, stage-based, scalable, serverless video processing framework that exploits intra-video parallelism to achieve low latency. enables developers program series of operations over content in modular, extensible manner. Programmers implement custom operations, ranging from simple transformations more complex computer vision tasks, pipeline specification language construct pipelines. then handles the underlying access, encoding and decoding, image across parallel...

Scale-out architectures supporting flexible, incremental scalability are common for computing and storage. However, the network remains last bastion of traditional scale-up approach, making it data center's weak link. Through UCSD Triton architecture, authors explore issues in managing as a single plug-and-play virtualizable fabric scalable to hundreds thousands ports petabits per second aggregate bandwidth.

The potential advantages of optics at high link speeds have led to significant interest in deploying optical switching technology data-center networks. Initial efforts focused on hybrid approaches that rely millisecond-scale circuit the core network, while maintaining flexibility electrical packet edge. Recent demonstrations microsecond-scale switches motivate considering for more dynamic traffic such as generated from a top-of-rack (ToR) switch. Based these trends, we propose prototype ToR,...

A range of new datacenter switch designs combine wireless or optical circuit technologies with electrical packet switching to deliver higher performance at lower cost than traditional packet-switched networks. These "hybrid" networks schedule large traffic demands via a high-rate circuits and remaining lower-rate, packet-switches. Achieving high utilization requires an efficient scheduling algorithm that can compute proper configurations balance across the switches. Recent proposals,...

Corundum is an open-source, FPGA-based prototyping platform for network interface development at up to 100 Gbps and beyond. The includes several core features enable real-time, high-line-rate operations including: a high-performance datapath, 10G/25G/100G Ethernet MACs, PCI Express gen 3, custom PCIe DMA engine, native high-precision IEEE 1588 PTP timestamping. A key feature extensible queue management that can support over 10,000 queues coupled with transmit schedulers, enabling...

This paper presents the design and implementation of an incrementally scalable architecture for middleboxes based on commodity servers operating systems. xOMB, eXtensible Open MiddleBox, employs general programmable network processing pipelines, with user-defined C++ modules responsible parsing, transforming, forwarding flows. We implement three pipelines in demonstrating good performance load balancing, protocol acceleration, application integration. In particular, our xOMB balancing switch...

Recent proposals have employed optical circuit switching (OCS) to reduce the cost of data center networks. However, relatively slow times (10--100 ms) assumed by these approaches, and accompanying latencies their control planes, has limited its use only largest networks with highly aggregated constrained workloads. As faster switch technologies become available, designing a plane capable supporting them becomes key challenge. In this paper, we design implement an OCS prototype in 11.5 us,...

We experimentally evaluate the network-level switching time of a functional 23-host prototype hybrid optical circuit-switched/electrical packet-switched network for datacenters called Mordia (Microsecond Optical Research Datacenter Interconnect Architecture). This uses standard electrical packet switch and an circuit-switched architecture based on wavelength-selective that has measured mean port-to-port reconfiguration 11.5 <formula formulatype="inline"...

Rate limiting is an important primitive for managing server network resources. Unfortunately, software-based rate suffers from limited accuracy and high CPU overhead, modern NICs only support a handful of limiters. We present SENIC, NIC design that can natively 10s thousands limiters--100x to 1000x the number available in today. The key idea host classifies packets, enqueues them per-class queues memory, specifies limits each traffic class. On NIC, SENIC maintains class metadata, computes...

FaaSnap is a VM snapshot-based platform that uses set of complementary optimizations to improve function cold-start performance for Function-as-a-Service (FaaS) applications. Compact loading files take better advantage prefetching. Per-region memory mapping tailors page fault handling depending on the contents different guest regions. Hierarchical overlapping memory-mapped regions simplify process. Concurrent paging allows start execution immediately, rather than pausing until working...

"Big Data" computing increasingly utilizes the MapReduce programming model for scalable processing of large data collections. Many jobs are I/O-bound, and so minimizing number I/O operations is critical to improving their performance. In this work, we present Themis, a implementation that reads writes records disk exactly twice, which minimum amount possible sets cannot fit in memory.

Recent proposals to build hybrid electrical (packet-switched) and optical (circuit switched) data center interconnects promise reduce the cost, complexity, energy requirements of very large networks. Supporting realistic traffic patterns, however, exposes a number unexpected difficult challenges actually deploying these systems "in wild." In this paper, we explore several challenges, uncovered during year experience using interconnects. We discuss both problems that must be addressed make...

While numerous studies have examined the macro-level behavior of traffic in data center networks---overall flow sizes, destination variability, and TCP burstiness---little information is available on at packet-level timescales. Whereas one might assume that flows from different applications fairly share link bandwidth, packets within a single are uniformly paced, reality more complex. To meet increasingly high rates 10 Gbps beyond, batching typically introduced across network stack---at...

We present TritonSort, a highly efficient, scalable sorting system. It is designed to process large datasets, and has been evaluated against as much 100 TB of input data spread across 832 disks in 52 nodes at rate 0.916 TB/min. When the annual Indy GraySort benchmark, TritonSort 60% better absolute performance over six times per-node efficiency previous record holder. In this paper, we describe hardware software architecture necessary operate level efficiency. Through careful management...

Using a compact (0.03 mm(2)) silicon-photonic bias-free thermo-optic cross-bar switch, we demonstrate microsecond-scale switching of twenty wavelength channels C-band wavelength-division multiplexed optical ring network, each carrying 10 Gbit/second data concurrently, with 15 mW electrical power consumption (no temperature control required). A convenient pulsed driving scheme is demonstrated and eye patterns bit-error rate measurements are shown. An algorithm developed to measure the...

Datacenters need networks that support both low-latency and high-bandwidth packet delivery to meet the stringent requirements of modern applications. We present Opera, a dynamic network delivers latency-sensitive traffic quickly by relying on multi-hop forwarding in same way as expander-graph-based approaches, but provides near-optimal bandwidth for bulk flows through direct over time-varying source-to-destination circuits. The key Opera's design is rapid deterministic reconfiguration...