1. |
- Hagsand, Olof, et al.
(author)
-
Control and forwarding plane separation on an opensource router
- 2011
-
Conference paper (peer-reviewed)abstract
- In previous work[1-3] it has been shown how open-source routers on new PC hardware allows for forwarding speeds of 10Gb/s and above. In this work we extend the applicability of the results by showing how the new 10Gb/s interface classification techniques can be used to separate packet forwarding from control plane operation.It is important to isolate the control-plane from forwarding load, since it makes routing protocol and management operation independent of forwarding load. It also increases the resilience against denial-ofservice attacks. In addition, it relates to the forwarding and control element separation proposed by the IETF ForCES work[4], where we use one CPU core as control element and the remaining cores as forwarding elements. Many new interface cards have chipsets with advanced classification capabilities motivated by advances in virtualization and multicore architectures. We have chosen to study the Intel 82599 10Gb/s controller[5] and the Linux ixgbe driver. The 82599 has several mechanisms to control packet classification, including Receiver Side Scaling (RSS)[6], Flow director, and N-tuple filters. Other interface cards on the market use generic TCAMs providing similar functionality. The approach we used was to implicitly configure the Flow director by outgoing control traffic, so that return flows aimed at the control plane were identified and could be directed to a designed control processor. Flows not destined to the control processor were load balanced among the remaining cores using RSS. We found this to be a simple and straight-forward approach, and we present results that verifies this method. However, we have seen some cases in overload scenarios where packet drops are made in hardware before classification which need to be further analyzed.During the project we also explored some of the hardware capabilites new buses (PCIe gen2). We discovered with optimal setting that we could transmit (DMA) 92 Gb/s using 1500 byte packet
|
|
2. |
- Hagsand, Olof, et al.
(author)
-
Measurement of IP forwarding performance on complex computer architectures
- 2011
-
In: Swedish National Computer Networking Workshop, SNCNW 2011.
-
Conference paper (peer-reviewed)abstract
- Open-source routers on new PC hardware allows for forwarding speeds of 10Gb/s and above. We present detailed performance measurements using Linux on two complex PC hardware platforms. Both platforms use PCIe gen2, dual I/O bridges and have support for non-uniform memory access (NUMA). The AMD platform uses four processors equipped with eight cores and four nodes of local memory. The Intel platform has two quad-core CPUs each with local memory. Packets being forwarded through a PC-based router can be separated into three steps: receive-dma, lookup, and transmitdma. Each step was studied individually. In particular, we studied how varying the CPU core and memory node effects the forwarding speeds. Our results show a large performance dependency of selecting CPU cores and memory nodes. In particular, DMA works best with memory nodes closest to the I/O bridge where the interface card is connected. Correspondingly, CPU access is most efficient on local memory. Consequently, choosing CPU core and memory nodes badly leads to a significant performance decrease.
|
|