Receive Descriptor Recycling for Small Packet High Speed Ethernet Traffic

Walravens, Cedric; Gaidioz, Benjamin

doi:10.1109/melcon.2006.1653329

Cited by 4 publications

(2 citation statements)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…RELATED WORK Our results and conclusions are similar to the discussion presented in the article [16]. In this article, the researchers obtained, for small packet size only, around 30% of the theoretical throughput for 1Gb/s.…”

Section: Splitting the Processing Into Independent Threadssupporting

confidence: 89%

Performance characteristics of multi-core and CMT servers

Nowostawski

McKenzie²,

Lipoth³

2007

2007 Australasian Telecommunication Networks and Applications Conference

View full text Add to dashboard Cite

Multi-core and chip-level multi-threading systems face serious challenges related to performance scaling. Their performance is dominated by I/0 bus crossings, cache misses and interrupts. In this article we analyse the performance characteristics of typical server systems utilising Sun Niagara T1, Intel Pentium D and AMD Opteron CPUs. Based on our results, we conclude that highly parallel CMT systems, such as the Sun T2000, have the potential to significantly benefit from their architecture -as long as their computational resources are used appropriately. Some of this allocation can be left to the operating system, but some must be tuned by careful hand design of the software, pipelines and dependencies between hardware strands.

show abstract

Section: Splitting the Processing Into Independent Threadssupporting

confidence: 89%

Performance characteristics of multi-core and CMT servers

Nowostawski

McKenzie²,

Lipoth³

2007

2007 Australasian Telecommunication Networks and Applications Conference

View full text Add to dashboard Cite

show abstract

“…Therefore, network virtualization technology can be used to efficiently configure and operate the network.However, open-source virtual switches used in network virtualization technology have a problem that the bandwidth of the network cannot be guaranteed. The reason for not guaranteeing the bandwidth of the network is because that buffer memory allocation / unassignment and interruption problems occur during the processing of large packets in the operating system kernel unit [2], [3]. The Data Plane Development Kit (DPDK) [4] is a technology developed to solve these problems.…”

Section: Introductionmentioning

confidence: 99%

Performance measurement using dpdk in software-defined local and wide area networks

Kim¹,

Kim²,

Kim³

et al. 2018

IJET

View full text Add to dashboard Cite

Background/Objectives: For data center-based enterprises, the use of data traffic continues to increase. As a result, the increasingly efficient processing of increasing data traffic is becoming a major issue. Effective processing of traffic can be performed through the addition of physical network equipment, but this requires considerable costs.Methods/Statistical analysis: To solve this problem, researches on server virtualization and network virtualization technology are being actively conducted. However, while network virtualization is a cost-effective solution, open-source virtual switches used in network virtualization technologies have the performance penalty because they handle packets at the kernel of server's operating system. As a result, the Intel has developed the Data Plane Development Kit (DPDK) technology to address the performance degradation of the operating system.Findings: In this paper, we propose the DPDK performance test environment based on Open Virtual Switch (OVS) to verify the stability and performance of the DPDK-oriented virtualization technology, in association with a new KREONET-S project which aims tothe softwarization of the Korea Research Environment Open Network (KREONET) infrastructure.Improvements/Applications: We perform conducted experiments on Software-Defined Local Area Network (LAN) and Wide Area Network (WAN) environment using KREONET-S show the result.

show abstract

Breaking up gigabit ethernet's VoIP bottlenecks

Walravens

Gaidioz²

2008

IEEE Potentials

View full text Add to dashboard Cite

GIGABIT ETHERNET IS A POPULAR BROADBAND technology for good reason. It not only can transfer large amounts of data through a large network at high speeds, but it also is a mature technology with an appealing performance to cost-of-ownership ratio. Unfortunately, though, gigabit Ethernet suffers a performance drop when it is handling traffic with small packets: it tends to lose them. This is a serious problem because small packets have become important in many real-time applications-in Voice over Internet Protocol (VoIP), for example, which carries voice conversations over the Internet. Every VoIP connection sends 50 packets over the network every second, packets that contain only 160 B or less. And since this small-packet, high-rate, real-time type of traffic is conveyed across the Internet, all Linux-based firewalls, routers, and Web caches suffer from the packet-loss problem too.Fortunately, there is a cure. The Ethernet controller's driver is the culprit. Its negligent memory management is responsible for the small-packet loss. A technique called receive descriptor recycling (RDR) provides a solution. By actively reusing the software descriptors needed to administer the traffic flow, RDR alleviates the controller bottleneck. An RDR-enabled controller can reduce small-packet loss by 40%.We have studied high-rate, smallpacket traffic in an Ethernet controller, as implemented in an Intel e1000 network interface card (NIC) Linux driver, which is widely used in high-end systems and servers. We have also evaluated the performance boost provided by the RDR technique. (For background, see "Researching Gigabit Ethernet Performance.") MEASURING NETWORK PERFORMANCEThe performance of a network can be measured in terms of packet rate, packet loss, response time, and throughput. In addition, important criteria for the network equipment are the efficient utilization of the central processing unit (CPU) and low resource requirements. When a network setup is optimized for high-rate small-packet traffic, inevitable limitations arise, e.g., CPU and resource requirements increase in order to accommodate high rates of small packets, as does the host bus utilization. Note that transmit performance is not affected by small-packet traffic to the same extent as receive performance. This asymmetry exists because the local host cannot usually overwhelm the Ethernet controller with outgoing traffic.

show abstract

Receive Descriptor Recycling for Small Packet High Speed Ethernet Traffic

Cited by 4 publications

References 6 publications

Performance characteristics of multi-core and CMT servers

Performance characteristics of multi-core and CMT servers

Performance measurement using dpdk in software-defined local and wide area networks

Breaking up gigabit ethernet's VoIP bottlenecks

Contact Info

Product

Resources

About