Optimizing burst-heavy datacenter workloads necessitates finegrained network control and visibility. We introduce CSIG, a protocol that delivers precise, multi-bit bottleneck congestion signals via a fixed-length Ethernet header. The architecture captures μsgranularity switch metrics, such as available bandwidth, and signals them to end-hosts using in-band, line-rate operations. We propose Fast Ramp-Up, a congestion control primitive that leverages these bottleneck signals to reduce median RPC latency by 20% and unclaimed bandwidth by 60% in production. Beyond transport-level performance, CSIG enables flow-aware observability by embedding μs-scale metrics into every packet, allowing individual application transfers to pinpoint their bottleneck location, such as the topology tier limiting their performance. CSIG thus transforms network telemetry from post-hoc correlation into a real time, context-aware capability. We demonstrate CSIG's broad deployability by validating it across five generations of commodity switch hardware (up to 102.4 Tbps), four NIC generations, and five transport stacks. Our design proves that a streamlined Layer 2 approach, focusing exclusively on the principal path bottleneck, provides transport-agnostic gains without requiring forklift hardware upgrades.

Gradient aggregation has long been identified as a major bottleneck in today's large-scale distributed machine learning training systems. One promising solution to mitigate such bottlenecks is gradient compression, directly reducing communicated gradient data volume. However, in practice, many gradient compression schemes do not achieve acceleration of the training process while also preserving accuracy.

An ISP's customers increasingly demand delivery of their traffic without congestion and with low latency. The ISP's topology, routing, and traffic engineering, often over multiple paths, together determine congestion and latency within its backbone. We first consider how to measure a topology's capacity to route traffic without congestion and with low latency. We introduce low-latency path diversity (LLPD), a metric that captures a topology's flexibility to accommodate traffic on alternative low-latency paths. We explore to what extent 116 real backbone topologies can, regardless of routing system, keep latency low when demand exceeds the shortest path's capacity. We find, perhaps surprisingly, that topologies with good LLPD are precisely those where routing schemes struggle to achieve low latency without congestion. We examine why these schemes perform poorly, and offer an existence proof that a practical routing scheme can achieve a topology's potential for congestion-free, low-delay routing. Finally we examine implications for the design of backbone topologies amenable to achieving high capacity and low delay.

Early in in the Internet's history, routing within a single provider's WAN centered on placing traffic on the shortest path. More recent traffic engineering efforts aim to reduce congestion and/or increase utilization within the status quo of greedy shortest-path first routing on a sparse topology. In this paper, we argue that this status quo of routing and topology is fundamentally at odds with placing traffic so as to minimize latency for users while avoiding congestion. We advocate instead provider backbone topologies that are more mesh-like, and hence better at providing multiple low-latency paths, and a routing system that directly considers latency minimization and congestion avoidance while dynamically placing traffic on multiple unequal-cost paths. We offer a research agenda for achieving this new low-latency approach to WAN topology design and routing.

As 802.11 wireless networks proliferate, interference becomes increasingly severe, particularly in dense, urban environments. These networks are usually operated by different users (e.g., tenants in apartments). In this paper, we develop techniques for mitigating interference between such loosely cooperating 802.11 MIMO APs and clients, which do not share a high-speed wired backplane or central controller. We propose CoOperative Power Allocation (COPA), an approach to concurrent wireless medium access that combines fine-grained, per-subcarrier power allocation, nulling, and multi-stream transmission to claim capacity that status-quo approaches cannot. Jointly turning these knobs allows COPA to allocate subcarriers to senders partially, rather than all-or-nothing, and to embrace a measure of interference when doing so increases capacity.

We present FUBAR, a system that reduces congestion and maximizes the utility of the entire network by installing new routes and changing the traffic load on existing ones. FUBAR works offline to periodically adjust the distribution of traffic on paths. It requires neither changes to end hosts nor precise prior knowledge of the traffic matrix. We demonstrate that even in the presence of traffic from all network devices to all other devices, FUBAR can optimize a real-world core-level network in a matter of minutes.

We consider the intra-AS route dissemination problem from first principles, and illustrate that when known route dissemination techniques propagate even a single external routing change, they can cause transient anomalies. These anomalies are not fundamental; they are artifacts of the order in which existing proposals disseminate routes. We show that carefully ordering route updates avoids transient looping and black holes. Perhaps surprisingly, this ordering may be enforced in a completely distributed fashion, while retaining familiar correctness, scalability, and convergence properties.

Dense 802.11 wireless networks present a pressing capacity challenge: users in proximity contend for limited unlicensed spectrum. Directional antennas promise increased capacity by improving the signal-to-interference-plus-noise ratio (SINR) at the receiver, potentially allowing successful decoding of packets at higher bit-rates. Many uses of directional antennas to date have directed high gain between two peers, thus maximizing the strength of the sender's signal reaching the receiver. But in an interference-rich environment, as in dense 802.11 deployments, directional antennas only truly come into their own when they explicitly null interference from competing concurrent senders. In this paper, we present Cone of Silence (CoS), a technique that leverages software-steerable directional antennas to improve the capacity of indoor 802.11 wireless networks by adaptively nulling interference. Using in situ signal strength measurements that account for the complex propagation environment, CoS derives custom antenna radiation patterns that maximize the strength of the signal arriving at an access point from a sender while nulling inteference from one or more concurrent interferers. CoS leverages multiple antennas, but requires only a single commodity 802.11 radio, thus avoiding the significant processing requirements of decoding multiple concurrent packets. Experiments in an indoor 802.11 deployment demonstrate that CoS improves throughput under interference.

Geographic techniques promise highly scalable any-to-any routing in wireless sensor networks. In one thread of research on geographic routing, researchers have explored robust, distributed graph planarization. Arguing that such planarization techniques have high overhead, researchers have more recently pursued a thread in which they propose precomputation of routing structures (e.g., hull trees and grids) to achieve low-overhead geographic routing.In this paper we introduce a third approach, LCR, that does not involve any precomputation of distributed routing structures, nor full a priori planarization. Instead, LCR removes non-planarities lazily only when they interfere with correct geographic routing. Lazy removal of link crossings results in an order of magnitude or more lower overhead than any previously proposed approach. Copyright 2006 ACM.

Large-scale distributed systems are hard to deploy, and distributed hash tables (DHTs) are no exception. To lower the barriers facing DHT-based applications, we have created a public DHT service called OpenDHT. Designing a DHT that can be widely shared, both among mutually untrusting clients and among a variety of applications, poses two distinct challenges. First, there must be adequate control over storage allocation so that greedy or malicious clients do not use more than their fair share. Second, the interface to the DHT should make it easy to write simple clients, yet be sufficiently general to meet a broad spectrum of application requirements. In this paper we describe our solutions to these design challenges. We also report our early deployment experience with OpenDHT and describe the variety of applications already using the system.

It is widely believed that content-signature-based intrusion detection systems (IDS) are easily evaded by polymorphic worms, which vary their payload on every infection attempt. In this paper, we present Polygraph, a signature generation system that successfully produces signatures that match polymorphic worms. Polygraph generates signatures that consist of multiple disjoint content substrings. In doing so, Polygraph leverages our insight that for a real-world exploit to function properly, multiple invariant substrings must often be present in all variants of a payload; these substrings typically correspond to protocol framing, return addresses, and in some cases, poorly obfuscated code. We contribute a definition of the polymorphic signature generation problem; propose classes of signature suited for matching polymorphic worm payloads; and present algorithms for automatic generation of signatures in these classes. Our evaluation of these algorithms on a range of polymorphic worms demonstrates that Polygraph produces signatures for polymorphic worms that exhibit low false negatives and false positives.

No abstract available.

TCP performs poorly on paths that reorder packets significantly, where it misinterprets out-of-order delivery as packet loss. The sender responds with a fast retransmit though no actual loss has occurred. These repeated false fast retransmits keep the sender's window small, and severely degrade the throughput it attains. Requiring nearly in-order delivery needlessly restricts and complicates Internet routing systems and routers. Such beneficial systems as multi-path routing and parallel packet switches are difficult to deploy in a way that preserves ordering. Toward a more reordering-tolerant Internet architecture, we present enhancements to TCP that improve the protocol's robustness to reordered and delayed packets. We extend the sender to detect and recover from false fast retransmits using DSACK information, and to avoid false fast retransmits proactively, by adaptively varying dupthresh. Our algorithm is the first that adaptively balances increasing dupthresh, to avoid false fast retransmits, and limiting the growth of dupthresh, to avoid unnecessary timeouts. Finally, we demonstrate that TCP's RTO estimator tolerates delayed packets poorly, and present enhancements to it that ensure it is sufficiently conservative, without using timestamps or additional TCP header hits. Our simulations show that these enhancements significantly improve TCP's performance over paths that reorder or delay packets.

We present Greedy Perimeter Stateless Routing (GPSR), a novel routing protocol for wireless datagram networks that uses the positions of routers and a packet's destination to make packet forwarding decisions. GPSR makes greedy forwarding decisions using only information about a router's immediate neighbors in the network topology. When a packet reaches a region where greedy forwarding is impossible, the algorithm recovers by routing around the perimeter of the region. By keeping state only about the local topology, GPSR scales better in per-router state than shortest-path and ad-hoc routing protocols as the number of network destinations increases. Under mobility's frequent topology changes, GPSR can use local topology information to find correct new routes quickly. We describe the GPSR protocol, and use extensive simulation of mobile wireless networks to compare its performance with that of Dynamic Source Routing. Our simulations demonstrate GPSR's scalability on densely deployed wireless networks.