Network bandwidth continues to scale rapidly, yet Internet data transmission performance remains constrained by the legacy 1500 B MTU. This small MTU translates high bandwidth into high packet rates that strain CPU processing at middleboxes and end hosts. While increasing the MTU could substantially improve performance, coordinating upgrades across arbitrary Internet paths is impractical.

This paper proposes a systematic approach to incrementally enabling large MTUs in the Internet. We demonstrate that increasing the MTU size significantly enhances the performance of both middleboxes and end hosts. To bridge MTU mismatches at network borders, we introduce PacketExpress gateway (PXGW), an MTU-translating gateway that dynamically adjusts packet sizes for cross-traffic. PXGW merges and splits TCP payloads on the fly and tunnels UDP packets, ensuring seamless adaptation. Also, we propose F-PMTUD, a new path MTU discovery algorithm that determines the path MTU within a single round-trip without relying on ICMP. Our preliminary evaluation shows that the PXGW prototype achieves 1.45 Tbps of packet forwarding throughput using only 8 CPU cores. After dynamic conversion, 94% of transmitted TCP packets are 9000 B jumbo frames, indicating that most flows were effectively converted into large segments, thereby demonstrating the system's efficiency and scalability. We also find that large-MTU packets, made available via PXGW, enhance end-host performance by up to 2.5X.

In modern server CPUs, the Last-Level Cache (LLC) serves not only as a victim cache for higher-level private caches but also as a buffer for low-latency DMA transfers between CPU cores and I/O devices through Direct Cache Access (DCA).However, prior work has shown that high-bandwidth network-I/O devices can rapidly flood the LLC with packets, often causing significant contention with co-running workloads.One step further, this work explores hidden microarchitectural properties of the Intel Xeon CPUs, uncovering two previously unrecognized LLC contentions triggered by emerging high-bandwidth I/O devices.Specifically, (C1) DMAwritten cache lines in LLC ways designated for DCA (referred to as DCA ways) are migrated to certain LLC ways (denoted as inclusive ways) when accessed by CPU cores, unexpectedly contending with non-I/O cache lines within the inclusive ways.In addition, (C2) high-bandwidth storage-I/O devices, which are increasingly common in datacenter servers, benefit little from DCA while contending with (latency-sensitive) network-I/O devices within DCA ways.To this end, we present A4, a runtime LLC management framework designed to alleviate both (C1) and (C2) among diverse co-running workloads, using a hidden knob and other hardware features implemented in those CPUs.Additionally, we demonstrate that A4 can also alleviate other previously known network-I/Odriven LLC contentions.Overall, it improves the performance of latency-sensitive, high-priority workloads by 51% without notably compromising that of low-priority workloads.

Cloud-based live video analytics with tight latency bound is gaining importance to support emerging applications such as UAVs and augmented reality. However, existing systems often struggle to meet stringent latency constraints under fluctuating network conditions with packet losses and late-arriving packets. We propose a loss-tolerant live video analytics system called Logan, which effectively accepts packet losses while maintaining high accuracy by utilizing the inherent resilience in DNNs. We design i) Codec-aware Inpainting, which accurately recovers the frame error from packet losses ii) Fast-Forward Recovery that prevents the remaining un-recovered error from propagating over future frames indefinitely. Our results show a 3× improvement (33.2%→99.9%) in SLO satisfaction rate compared to the reliable transmission scheme with <1% accuracy drop under a 5% packet loss rate.

The Maximum Transmission Unit (MTU) refers to the largest packet size that can be transferred on a particular layer-3 network. As the dominance of Ethernet prevails, the "de-facto" standard MTU of 1500B has become universal in the wide-area networks. Unfortunately, the current MTU size overly limits the transmission performance especially when the underlying link speed rapidly increases while the CPU advancement stagnates.

Transport Layer Security (TLS) has become a key building block for private network communication in modern Internet. While recent advancement of CPU has substantially improved the data encryption performance, TLS key exchange still remains the bottleneck for short-lived transactions. Dedicated hardware crypto accelerators promise good performance, but they often require invasive modification of the application due to its inherent architecture of asynchronous processing.

Determinantal point processes (DPPs) are popular probabilistic models that arise in many machine learning tasks, where distributions of diverse sets are characterized by determinants of their features. In this paper, we develop fast algorithms to find the most likely configuration (MAP) of large-scale DPPs, which is NP-hard in general. Due to the submodular nature of the MAP objective, greedy algorithms have been used with empirical success. Greedy implementations require computation of log-determinants, matrix inverses or solving linear systems at each iteration. We present faster implementations of the greedy algorithms by utilizing the orthogonal benefits of two log-determinant approximation schemes: (a) first-order expansions to the matrix log-determinant function and (b) high-order expansions to the scalar log function with stochastic trace estimators. In our experiments, our algorithms are orders of magnitude faster than their competitors, while sacrificing marginal accuracy.

Ensemble methods are arguably the most trustworthy techniques for boosting the performance of machine learning models. Popular independent ensembles (IE) relying on naive averaging/voting scheme have been of typical choice for most applications involving deep neural networks, but they do not consider advanced collaboration among ensemble models. In this paper, we propose new ensemble methods specialized for deep neural networks, called confident multiple choice learning (CMCL): it is a variant of multiple choice learning (MCL) via addressing its overconfidence issue.In particular, the proposed major components of CMCL beyond the original MCL scheme are (i) new loss, i.e., confident oracle loss, (ii) new architecture, i.e., feature sharing and (iii) new training method, i.e., stochastic labeling. We demonstrate the effect of CMCL via experiments on the image classification on CIFAR and SVHN, and the foreground-background segmentation on the iCoseg. In particular, CMCL using 5 residual networks provides 14.05\% and 6.60\% relative reductions in the top-1 error rates from the corresponding IE scheme for the classification task on CIFAR and SVHN, respectively.

Delay-tolerant Wi-Fi offloading is known to improve overall mobile network bandwidth at low delay and low cost. Yet, in reality, we rarely find mobile apps that fully support opportunistic Wi-Fi access. This is mainly because it is still challenging to develop delay-tolerant mobile apps due to the complexity of handling network disruptions and delays. In this paper, we present Cedos, a practical delay-tolerant mobile network access architecture in which one can easily build a mobile app. Cedos consists of three components. First, it provides a familiar socket API whose semantics conforms to TCP, while the underlying protocol, D2TP, transparently handles network disruptions and delays in mobility. Second, Cedos allows the developers to explicitly exploit delays in mobile apps. App developers can express maximum user-specified delays in content download or use the API for real-time buffer management at opportunistic Wi-Fi usage. Third, for backward compatibility to existing TCP-based servers, Cedos provides D2Prox, a protocol-translation Web proxy. D2Prox allows intermittent connections on the mobile device side, but correctly translates Web transactions with traditional TCP servers. We demonstrate the practicality of Cedos by porting mobile Firefox and VLC video streaming client to using the API. We also implement delay/disruption-tolerant podcast client and run a field study with 50 people for eight weeks. We find that up to 92.4% of the podcast traffic is offloaded to Wi-Fi, and one can watch a streaming video in a moving train while offloading 48% of the content to Wi-Fi without a single pause.

In this work, we present a highly scalable network intrusion detection system on many-core processors. To maximize the NIDS performance, we take advantage of the underlying hardware and adhere to four design principles: shared-nothing architecture, computation offloading, lightweight data structure, and flow offloading. Through the experimental results, we find that our design choices can significantly improve the NIDS performance (79 Gbps with 1514B synthetic packets). We believe that our design decisions can be easily extended to other many-core processors and programmable NICs.

Delay-tolerant Wi-Fi offloading is known to improve overall mobile network bandwidth at low delay and low cost. Yet, in reality, we rarely find mobile apps that fully support opportunistic Wi-Fi access. This is mainly because it is still challenging to develop delay-tolerant mobile apps due to the complexity of handling network disruptions and delays.

No abstract available.

Caching similar videos transparently in a network is a cost-effective solution that potentially reduces redundant data transfers. Recent study shows that network redundancy elimination (NRE) on the content level could produce high bandwidth savings in ISPs. However, we find that blindly employing existing NRE techniques to video contents could lead to suboptimal redundancy suppression rates. This is because (a) randomness in the video encoding process could produce completely different binaries even when they deal with seemingly identical video clips and (b) existing NRE chunking schemes incur high overheads since they do not utilize the underlying video format.

Vehicular networks have traditionally focused on the real-time delivery of critical control messages for safe car operation. Unfortunately, the real-time requirements often cripple the development of flexible car applications by tying the application network stack to underlying physical networks. While popular real-time vehicular networks guarantee the timely delivery of prioritized messages, they often lack in bandwidth and flexibility, which limits the range of car network applications. In this work, we explore the idea of replacing the current vehicular network with standard switched Ethernet, the most popular LAN technology in computer networks. Ethernet is attractive in providing high bandwidth at a low cost with easy and flexible configuration. The most challenging part is to guarantee the real-time delivery of mission-critical messages. We first show that the soft message delivery latency of 10s to 100s milliseconds can be easily met in 100 Mbps switched Ethernet despite coexistence of high-bandwidth network applications. For meeting the hard delivery latency on the order of 100 microseconds for critical control messages, we propose limiting the path MTU to the destination node with priority queuing from IEEE 802.1Q. Our simulation shows that we can satisfy 100 microseconds of latency even in a rich set of vehicular applications without any modification of the application network stack.

Recent popularity of smartphones drives rapid growth in the demand for cellular network bandwidth. Unfortunately, due to the centralized architecture of cellular networks, increasing the physical backhaul bandwidth is challenging. While content caching in the cellular network could be beneficial, relatively few characteristics of the cellular traffic is known to come up with a highly-effetive caching strategy. In this work, we provide insight into flow and content-level characteristics of modern 3G traffic at a large cellular ISP in South Korea. We first develop a scalable deep flow inspection (DFI) system that can manage hundreds of thousands of concurrent TCP flows on a commodity multicore server. Our DFI system collects various HTTP/TCP-level statistics and produces logs for analyzing the effectiveness of conventional Web caching, prefix-based Web caching, and TCP-level redundancy elimination (RE) without a single packet drop at a 10~Gbps link. Our week-long measurements of over 370 TBs of the 3G traffic reveal that standard Web caching can reduce download bandwidth consumption up to 27.1% while simple TCP-level RE can save the bandwidth consumption up to 42.0% with a cache of 512~GB of RAM. We also find that applying TCP-level RE on the largest 9.4% flows eliminates 68.4% of the total redundancy. Most of the redundancy (52.1%~58.9%) comes from serving the same HTTP objects while the contribution by aliased URLs is up to 38.9%.

As high-speed networks are becoming commonplace, it is increasingly challenging to prevent the attack attempts at the edge of the Internet. While many high-performance intrusion detection systems (IDSes) employ dedicated network processors or special memory to meet the demanding performance requirements, it often increases the cost and limits functional flexibility. In contrast, existing software-based IDS stacks fail to achieve a high throughput despite modern hardware innovations such as multicore CPUs, manycore GPUs, and 10 Gbps network cards that support multiple hardware queues.

Commodity-hardware technology has advanced in great leaps in terms of CPU, memory, and I/O bus speeds. Benefiting from the hardware innovation, recent software routers on commodity PC now report about 10 Gbps in packet routing. In this paper we map out expected hurdles and projected speed-ups to reach 100 Gbps in packet routing on a single commodity PC. With careful measurements, we identify two notable bottlenecks for our goal: CPU cycles and I/O bandwidth. For the former, we propose reducing per-packet processing overhead with software-level optimizations and buying extra computing power with GPUs. To improve the I/O bandwidth, we suggest scaling the performance of I/O hubs that limits packet routing speed to well before 50 Gbps.

We present PacketShader, a high-performance software router framework for general packet processing with Graphics Processing Unit (GPU) acceleration. PacketShader exploits the massively-parallel processing power of GPU to address the CPU bottleneck in current software routers. Combined with our high-performance packet I/O engine, PacketShader outperforms existing software routers by more than a factor of four, forwarding 64B IPv4 packets at 39 Gbps on a single commodity PC. We have implemented IPv4 and IPv6 forwarding, OpenFlow switching, and IPsec tunneling to demonstrate the flexibility and performance advantage of PacketShader. The evaluation results show that GPU brings significantly higher throughput over the CPU-only implementation, confirming the effectiveness of GPU for computation and memory-intensive operations in packet processing.

SSL/TLS is a standard protocol for secure Internet communication. Despite its great success, today's SSL deployment is largely limited to security-critical domains. The low adoption rate of SSL is mainly due to high computation overhead on the server side.