As 5G is getting closer to being commercially available, base stations processing this traffic must be improved to be able to handle the increase in traffic and demand for lower latencies. By utilizing the hardware smarter, the processing of data can be accelerated in, for example, the forwarding plane where baseband and encryption are common tasks. With this in mind, systems with integrated GPUs becomes interesting for their additional processing power and lack of need for PCIe buses.This thesis aims to implement the DPDK framework on the Nvidia Jetson Xavier system and investigate if a scheduler based on the theoretical properties of each platform is better than a self-exploring machine learning scheduler based on packet latency and throughput, and how they stand against a simple round-robin scheduler. It will also examine if it is more beneficial to have a more flexible scheduler with more overhead than a more static scheduler with less overhead. The conclusion drawn from this is that there are a number of challenges for processing and scheduling on an integrated system. Effective batch aggregation during low traffic rates and how different processes affect each other became the main challenges.

No description available.

Data Plane Development Kit is a highly used library for creating network applications that can be run on all hardware. Data Plane Development Kit has a component called Service Cores, which allows the main applications to create services that will run independently. These services are manually mapped to specific CPU cores, and are scheduled in a round-robin method. Because of the manual mapping, and the scheduling, the different load for each service can impact the start time for each service. By having services not run when supposed to, the throughput will degrade. In this thesis, we investigate and try to solve the issue by implementing a basic load balancer into the Service Core component. Our results show that an basic load balancer, that will balance upon reaching a CPU upper threshold, will increase the throughput of services while decreasing the delay between each service run.

Cloud-based video processing is an area depending heavily on the hardware’s ability to process huge amounts of packets. Nowadays, we can observe industry drifting away from commonly used FPGAs in lieu of a more flexible software approach. Docker container has been considered a promising technology for constructing video streaming pipelines as it provides a fast and easy way to package and distribute software. Recent developments in the Network Function Virtualization field showed that fast packet processing frameworks like Intel Data Plane Development Kit (DPDK) have a potential to improve the performance of network function chains. This technology could be an enabler for software video processing to approach hardware solutions, yet it is still in quite an early stage and generates many questions about usage, deployment, and performance. This thesis shows that it is possible to build packet processing pipelines using DPDK, running dozens of video processing microservices simultaneously on a single machine. The project implementation was evaluated in terms of latency and throughput and behaviour of co-running applications on a single CPU core was modelled. / Molntjänster Inom området videobearbetning är starkt beroende av hårdvarans förmåga att kontinuerligt bearbeta mycket stora mängder paket. Idag har det blivit vanligt inom professionell användning att välja bort tidigare vanliga FPGA-baserade lösningar, till förmån för mer flexibla mjukvarubaserade lösningar. Docker containers har setts som en lovande teknologi för att konstruera pipelines för strömmande video, då de erbjuder ett snabbt och enkelt sätt att paketera och distribuera mjukvara. Utvecklingen inom virtualisering av nätverksfunktioner (NFV) har visat att ramverk för snabb paketprocessning, såsom Intel DPDK, har potential att förbättra prestandan hos kedjor av nätverksfunktioner. Denna teknologi gör det möjligt för mjukvarubaserad videobehandling att hävda sig i jämförense med hårdvarubaserade varianter. Den är dock relativt ny och oprövad och det kvarstår många öppna frågor om användning, driftsättning och prestanda. Detta examensarbete visar att det är möjligt att bygga paketbearbetande pipelines med DPDK, som kör dussintals nätverksfunktioner samtidigt på en maskin. En implementation har konstruerats och utvärderats med fokus på latens och flöde, och beteendemönster för applikationer som kör samtidigt på samma CPU har modellerats.

V dnešnom svete zrýchľujúcej sa sieťovej prevádzky je potrebné držať krok v jej monitorovaní . Dostatočný prehľad o dianí v sieti dokáže zabrániť rozličným útokom na ciele nachádzajúce sa v nej . S tým nám pomáhajú systémy IDS, ktoré upozorňujú na udalosti nájdené v analyzovanej prevádzke . Pre túto prácu bol vybraný systém Suricata . Cieľom práce je vyladiť nastavenia systému Suricata s rozhraním AF_PACKET pre optimálnu výkonnosť a následne navrhnúť a implementovať optimalizáciu Suricaty . Výsledky z meraní AF_PACKET majú slúžiť ako základ pre porovnanie s navrhnutým vylepšením . Navrhovaná optimalizácia implementuje nové rozhranie založené na projekte Data Plane Development Kit ( DPDK ). DPDK je schopné akcelerovať príjem paketov a preto sa predpokladá , že zvýši výkon Suricaty . Zhodnotenie výsledkov a porovnanie rozhraní AF_PACKET a DPDK je možné nájsť na konci diplomovej práce .

The world of computer networks is constantly evolving towards more efficient algorithms and more effective ways of using hardware resources. One of these evolutions is the eXpress Data Path (XDP). XDP is an interrupt based data path in the Linux kernel. XDP uses JIT-compiled programs in a virtual machine in a device driver context. Through XDP, fast packet processing is possible while still keeping the functionality of the Linux kernel intact. Therefore, this thesis aims to illuminate possible use cases for XDP in 5G simulators, as this real-world application of XDP is of interest. Specifically, use-cases where there is a need for fast packet processing. This thesis evaluates the use-cases using a performance evaluation of XDP and a literary review of 5G simulators, XDP, and technologies relating to XDP. This evaluation indicates that XDP is a candidate for packet processing in 5G simulators, specifically when compared to what performance is possible currently. This thesis argues from the performance evaluation and the literary review that XDP can be used for small programs, preferably data ingestion, in 5G simulators.

One recent innovation in the networking industry, is the concept of Network FunctionVirtualization (NFV). NFV is based on a networking paradigm in which network functions,which have typically been implemented in the form of dedicated hardware appliances in thepast, are implemented in software and deployed on commodity hardware using modernvirtualization techniques. While the most common approach is to place each virtual networkfunction in a virtual machine - using hardware-level virtualization – the growing influenceand popularity of Docker and other container-based solutions has naturally led to the idea ofcontainerized deployments. This is a promising concept, as containers (or operating systemlevel virtualization) can offer a flexible and lightweight alternative to hardware-levelvirtualization, with the ability to use the resources of the host directly. The main problem withthis concept, is the fact that the default behavior of Docker and similar technologies is to relyon the networking stack of the host, which typically isn’t performant enough to handle theperformance requirements associated with NFV. In this dissertation, an attempt is made toevaluate the feasibility of using userspace networking to accelerate the network performanceof Docker containers, bypassing the standard Linux networking stack by moving the packetprocessing into userspace.

The performance of the virtual switch Open vSwitch (OVS) is insufficient to satisfy the current requirements for link bandwidth of the server connections. There is an effort to accelerate the OVS both in the software and in the hardware by offloading the datapath to the smart network interface cards. In this work the COMBO card for 100G Ethernet developed by CESNET is used to accelerate the OVS. The suggested solution utilizes the firmware for FPGA generated from the definition in the P4 language to classify the packets in the card and DPDK for the data transfers and offloading the classification rules into the card. Forwarding of one flow with the shortest frames from physical to physical interface using one CPU core reaches forwarding rate of 11.2 Mp/s (10 times more than the standard OVS) with classification in the card and 5.9 Mp/s without classification in the card.

This thesis investigates the network performance of DPDK, io_uring, and the standard Linux network stack in terms of latency, packet loss, throughput, and packet rate. DPDK is widely adopted in the telecommunication industry and has been for several years. io_uringis a new Linux system call interface, as of this writing, used for asynchronous I/O operations and could potentially offer better network performance than the standard Linux network stack. The analysis involves benchmarking of DPDK, io_uring and Linux network stack both with a continuous flow of data traffic and bursty traffic to see how the applications are able to handle different kinds of data traffic conditions. The Linux kernel network parameters were tuned in order to investigate what kind of performance increase can be achieved for an application using io_uring to gain a better understanding of howio_uring compares to DPDK. From the analysis, it was concluded that DPDK had the best overall performance when considering packet loss, throughput, latency, and packet rate.io_uring performed better than the Linux network stack in terms of throughput, packet loss, and packet rate but was not considerably better in terms of latency. When comparing the latency of bursty traffic for DPDK, io_uring, and the standard Linux network stack, it was shown that the kernel-based alternatives were better able at handling traffic over their maximum performance than DPDK, thus it may offer better scalability. After tuning the Linux kernel network parameters it was concluded that tuning may impact the performance network metrics of an application using io_uring. We conclude that io_uring shows some promising results compared to the standard Linux network stack, but is not yet able to compare with DPDK in network performance. The development of io_uring is ongoing and it may improve further in the future.

Recently, we have seen an evolution of programmable network devices, where it is possible to customize packet processing inside the data plane at an unprecedented level. This is in contrast to traditional approaches, where networking device functionality is fixed and defined by the ASIC and customers need to wait possibly years before the vendors release new versions that add features required by customers. The vendors in the industry have adapted and the focus has shifted to offering new types of network devices, such as the SmartNIC, IPU, and DPU. Another major paradigm shift in the networking area is the shift towards protocols that encrypt parts of headers and contents of packets such as QUIC. Also, many devices such as smart phones have support for multiple access networks, which requires efficient multipath protocols to leverage the capabilities of multiple networks at the same time. However, when using protocols inside the network that requires encryption such as QUIC or multipath QUIC, packet processing operations for the en/decryption process are very resource intensive. Consequently, network vendors and operators are in need to accelerate and offload crypto operations to dedicated hardware in order to free CPU cycles for business critical operations. Therefore, the aim of this study is to investigate how multipath QUIC can be offloaded or hardware accelerated in order to reduce the CPU utilization on the server. Our contributions are an evaluation of frameworks, programming languages and hardware devices in terms of crypto offloading functionality. Two packet processing offloading prototypes were designed using the DPDK framework and the programming language P4. The design using DPDK was implemented and evaluated on a BlueField 2 DPU. The offloading prototype handles a major part of the packet processing and the crypto operations in order to reduce the load of the user application running on the host. A evaluation show that the throughput when using larger keys are only slightly decreased. The evaluation gives important insights in the need of crypto engines and/or CPUs with high performance when offloading.