Introduction

  • The demand for bandwidth between servers inside a data center is rapidly increasing
    • High bisection bandwidth and high availability
    • Efficient use of multiple paths between pair of end hosts
  • The controller needs to use switch hardware resources or multiple flow tables efficiently for constructing a scalable network
  • OpenFlow is not sufficient for fast failover for high availability in the multipath
  • DomainFlow: a practical flow management using OpenFlow for commodity switches

Background

TCAM for Wild-carded Matching Rule

  • Ethernet switches uses TCAMs to implement wild-carded OpenFlow rules
  • TCAM consumes much higher area and power per entry than SRAM
    • On-chip TCAM sizes are typically limited to a few thousand entries

Overhead for Flow Level Management

  • Efficient multipath usage is possible by selecting path for every flow
  • The number of matching rules increases when we specify each flow for multipath distribution
  • Specify each flow for multipath distribution increases the number of matching rules
    • OpenFlow requires wild-carded matching rules for traffic distribution
  • OpenFlow-based path selection is limited due to the switch hardware resources
  • OpenFlow provides a group action to select path from multiple output ports
    • The distribution algorithm is not defined in the specification

Fast Failover for High Availability

  • OpenFlow defines a fast failover mechanism for the controller to specify the alternate port
    • Enabling the switch to change forwarding without requiring a round trip to the controller
    • However, it does not work with multiple switches in the multipath

DomainFlow

DomainFlow Flow Model

  • The flow model consists of two domains
    • Controller configures the boundary (turning point) on a per flow basis
  • In Domain 1
    • A flow is controlled with a wild-carded matching table (WMT)
    • For path selection
    • The flow is handled by the WMT if destination lookup failure (DLF) occurs in the EMT
  • In Domain 2
    • A flow is controlled with an exact matching table (EMT)
    • For unicast, multicast, broadcast packets
    • If DLF occurs, handled by the controller
  • Using proactive mode to reduce the flow setup overhead

Figure 1

DomainFlow Use Case

  • DomainFlow can be used in various types of applications
    • Path selection in multipath networks with tree structure
    • End-host mode to represent servers
    • Forwarding of FCoE (Fibre Channel over Ethernet) traffic to FCF (FCoE Forwarder)
  • Domain 1 is defined in which no destination exists at the server-facing ports of a switch
  • Domain 2 is defined in which a destination exists at the server-facing ports of a switch
  • The upward flow from a source server is controlled with the WMT towards the turning point
  • The downward flow is controlled with the EMT towards a destination server

Figure 2

DomainFlow Flow Table Structure

  • A commodity switch normally has both an EMT and WMT
    • The EMT is implemented as a forwarding database (FDB) using SRAM
    • The WMT is implemented as an access control list (ACL) using TCAM
  • The priority of EMT is higher than that of WMT
  • Use bit maksing to further reduce the number of WMT entries

Figure 3

Fast Failover in Multipath

  • The controller proactively configures a group type of Fast Failover
  • This mechanism works only when a switch detects a link failure on the directly connected link
  • Notify other affected switches about the link failure
  • Appropriately change forwarding without intervention of the controller

Figure 5

Application

Applied the DomainFlow method to multipath flow management in the VXLAN overlay network to enable network virtualization

VXLAN Packet Encapsulation

  • VXLAN tunnel end point (VTEP) originates and terminates VXLAN tunnels
  • Additional headers are Outer Ethernet, Outer IP, Outer UDP, and VXLAN
  • The VXLAN header includes 24 bits of VXLAN network identifier (VNI)

Figure 6

Multipath Traffic Distribution

  • Traffic from/to virtual machines is aggregated/disaggregated at a VTEP
  • It is important to distribute traffic in multipath networks reflecting original payloads for high performance
    • The UDP source port of the outer header to be a hash of the inner Ethernet frame’s header

DomainFlow Flow Management

  • OpenFlow Vendor Action is of the form:

Figure 9

Protype & Evaluation

Switch Prototype

  • Based on OpenFlow 1.0
  • Running an OpenFlow agent connected to the switch
    • Communicates with the controller over secure channel
    • Configures the switch using the enhanced CLI

Figure 10

Preliminary Evaluation Results

  • Can be implemented using a commodity switch
  • Small number of flow table entries

Table 2

Flow Table Efficiency

  • Traditional method installed matching rules in the TCAM
  • DomainFlow use FDB and ACL to reduce the total number of flow entries

Figure 11

Conclusion

  • Multipath flow management is possible with a small number of flow entries
    • Bit masking and wild-carded matching in Domain 1
    • Exact matching in Domain 2
  • Fast failover using OpenFlow for commodity switches
  • Applied a prototype of DomainFlow
    • Reducing the total number of flow entries to

Questions

  • Use bit masking on what to reduce the WMT entries
  • Wild-carded matching
  • Overlay network
  • OpenFlow Fast Failover mechanism
  • Vendor action for path selection

Notes

VXLAN

  • VNI stands for “VXLAN Network Identifier” (or VXLAN Segment ID)
  • The term “VXLAN segment” is used interchangeably with the term “VXLAN overlay network”
  • The end point of the tunnel (VXLAN Tunnel End Point or VTEP) is located within the hypervisor on the server that hosts the VM
  • VMs in different VXLAN overlay networks cannot communicate with each other

TCAM

  • TCAM 以內容查找地址,與一般 memory 不同
  • 一般 memory 中每個 bit 只能表示兩個值:0 與 1。而 TCAM 每個 bit 可以表示三個值:0、1 與 X,X 表示 don’t care
  • TCAM 的一條 entry 在物理上是用兩條 entry 所組成,一條放 data、另一條存放相應的 mask。因此可實現 0/1/X 的表示
  • TCAM 可支持各種 lookup key 任意組合查找,任意 mask 掉任何不關心的字段
  • 一條 TCAM entry 至少抵得上五條 DRAM/SRAM entry,如果晶片內部放了太大的 TCAM 會導致晶片成本和功耗直線上升

References