In recent years, we have witnessed the emergence of high speed packet I/O frameworks, bringing unprecedented network performance to userspace. Using Click modular router, rst review and quantitatively compare several such showing their superiority kernel-based forwarding. We then reconsider issue software processing, in context modern commodity hardware with multi-queues, multi-core processors non-uniform memory access. Through a combination existing techniques improvements our own, derive...

In recent years, we have witnessed the emergence of high speed packet I/O frameworks, bringing unprecedented network performance to userspace. Using Click modular router, first review and quantitatively compare several such showing their superiority kernel-based forwarding. We then reconsider issue software processing, in context modern commodity hardware with multi-queues, multi-core processors non-uniform memory access. Through a combination existing techniques improvements our own, derive...

We present PacketMill, a system for optimizing software packet processing, which (i) introduces new model to efficiently manage metadata and (ii) employs code-optimization techniques better utilize commodity hardware. PacketMill grinds the whole processing stack, from high-level network function configuration file low-level userspace (specifically DPDK) drivers, mitigate inefficiencies produce customized binary given function. Our evaluation results show that increases throughput (up 36.4...

While the current literature typically focuses on load-balancing among multiple servers, in this paper, we demonstrate importance of within a single machine (potentially with hundreds CPU cores). In context, propose new technique (RSS++) that dynamically modifies receive side scaling (RSS) indirection table to spread load across cores more optimal way. RSS++ incurs up 14x lower 95th percentile tail latency and orders magnitude fewer packet drops compared RSS under high utilization. allows...

Large service providers use load balancers to dispatch millions of incoming connections per second towards thousands servers. There are two basic yet critical requirements for a balancer: <i>uniform distribution</i> the across servers, which requires support advanced balancing mechanisms, and <i>per-connection-consistency</i> (PCC), i.e, ability map packets belonging same connection server even in presence changes number active servers balancers. Yet, simultaneously meeting these has been an...

As network speeds have increased to over 100 Gbps, operators and researchers lost the ability easily ask complex questions of reassembled parsed traffic. In this paper, we introduce Retina, a software framework that lets users analyze Gbps real-world traffic on single server with no specialized hardware. Retina supports running arbitrary user-defined analysis functions wide variety extensible data representations ranging from raw packets application-layer handshakes. We novel filtering...

To cope with the growing performance needs of appliances in datacenters or network edge, current middle-box functionalities such as firewalls, NAT, DPI, content-aware optimizers load-balancers are often implemented on multiple (perhaps virtual) machines. In this work, we design a system able to run pipeline VNFs high level parallelism handle many flows. We provide user facilities define traffic class interest for VNF, definition session group packets TCP 4-tuples, and amount space per...

The Internet use IP addresses to identify and locate network interfaces of connected devices. IPv4 was introduced more than 40 years ago specifies 32-bit addresses. As the grew, available eventually became exhausted ten ago. IETF designed IPv6 with a much larger addressing space consisting 128-bit addresses, pushing back exhaustion problem further in future. In this paper, we argue that large allows reconsidering how are used enables improving, simplifying scaling Internet. By revisiting...

Several implementations of the QUIC protocol exist. Unfortunately, they generally lack behind TCP ones in terms performance as stacks have been undergoing years optimizations. In this work, we propose picoquic-dpdk, a modified version picoquic that bypasses Linux kernel networking stack using DPDK library, improving throughput by 3x factor. We compare our implementation against several and TCP+TLS demonstrate it outperforms all tested matches even with common

Deployment of 100Gigabit Ethernet (GbE) links challenges the packet processing limits commodity hardware used for Network Functions Virtualization (NFV). Moreover, realizing chained network functions (i.e., service chains) necessitates use multiple CPU cores, or even servers, to process packets from such high speed links. Our system Metron jointly exploits underlying and servers’ resources: ( i ) offload part logic network, ii by using smart tagging setup exploit affinity traffic classes,...

Today, datacenter operators deploy Load-balancers (LBs) to efficiently utilize server resources, but must over-provision resources (by up 30%) because of load imbalances and the desire bound tail service latency. We posit one reasons for these is lack per-core statistics in existing LBs. As a first step, we designed CrossRSS, CPU core-aware LB that dynamically assigns incoming connections least loaded cores pool. CrossRSS leverages knowledge dispatching by each server's Network Interface...

<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Network functions</i> such as firewalls, NAT, DPI, content-aware optimizers, and load-balancers are increasingly realized software to reduce costs enable outsourcing. To meet performance requirements these xmlns:xlink="http://www.w3.org/1999/xlink">virtual</i> network functions (VNFs) often bypass the kernel use their own user-space networking stack. A naïve realization of a chain VNFs will...

The rise of commodity servers equipped with high-speed network interface cards poses increasing demands on the efficient implementation connection tracking, i.e., task associating identifier an incoming packet to state stored for that connection. In this work, we thoroughly investigate and compare performance obtainable by different implementations tracking using real traffic traces. Based a load balancer use case, our results show is expensive operation, achieving at most 24 Gbps single...