differentiated congestion management of data traffic for data center ethernet

Differentiated Congestion Management of Data Traffic for Data center EthernetB99705025 資管三 陳育旋

author

Shuo Fang, Student Member, IEEE, Chuan Heng Foh, Senior Member, IEEE,

and Khin Mi Mi Aung, Senior Member, IEEE

Introduction

Fibre Channel over Ethernet (FCoE) -- use Ethernet technology to carry Fibre Channel -- Fibre Channel are encapsulated in Ethernet to be transmitted

FCoE has some benefit -- reduce power consumption for I/O operation -- eliminate redundancy in the network -- I/O consolidation among Local Area Networks(LANs) and Storage Area Networks(SANs)

SANs (reference)

• Storage Area Networks• 將許多儲存裝置從區域網路獨立出來成為另一個網路• 透過光纖通道﹙ Fibre Channel﹚ 與伺服主 機做連結• 具備快速備份與災難復原能力• 但成本較高、且建置費時 http://goo.gl/TxKBj

• However…. -- Fibre Channel (FC) is designed to achieve high speed

lossless packet transportation -- Ethernet has no control on the traffic congestion

Introduction

How to consolidate?Differentiated congestion control in Ethernet!

outline

• Introduction• Ethernet Congestion Management• Differentiated Congestion Control• Analysis (experiment)• conclusion

outline

• Introduction• Ethernet Congestion Management• Differentiated Congestion Control• Analysis (experiment)• conclusion

Ethernet Congestion Management

• Implementing a particular ECN at switches and rate limiter at host

Ethernet Congestion Management

Qoff is the offset of the current buffer utilization with respect to Qeq (+ or - )

Qdelta is the change in length of the queue since the last sampled frame

• In ECM, each rate limiter implements a variation of AIMD for its rate adjustment.

• A rate limiter periodically increase its sending rate

• Fb = feedback signal

Ethernet Congestion Management

outline

• Introduction• Ethernet Congestion Management• Differentiated Congestion Control• Analysis (experiment)• conclusion

Differentiated Congestion Control

• we propose using different AIMD parameter sets for the rate limiters to achieve congestion control differentiation

• Queue Management in Congestion Point• Rate Limiters in the Reaction Point

Differentiated Congestion Control

• Queue Management in Congestion Point• Rate Limiters in the Reaction Point

Differentiated Congestion Control

• A congestion point (CP) features a queue management

• which is mainly responsible for congestion detection, congestion , notification and packet drop policy.

• when the buffer utilization exceeds Tsc , all LAN traffic, being the low priority traffic,

will be dropped.

Queue Management in Congestion Point

• Queue Management in Congestion Point• Rate Limiters in the Reaction Point

Differentiated Congestion Control

Rate Limiters in the Reaction Point

• A rate limiter regulates the traffic flow by controlling the transmission rate using an AIMD based rate adjustment operation

• the rate limiter maintains a variable called congestion window to regulate the transmission rate

• The value of congestion window is initialized to be one. This value increases linearly over a predefined constant time interval called a slot.

Rate Limiters in the Reaction Point

outline

• Introduction• Ethernet Congestion Management• Differentiated Congestion Control• Analysis (experiment)• conclusion

Analysis(experiment)

Background traffic

SAN(a1,b1)

LAN(a2,b2)

destination

destination

destinationBuffer(bottleneck)

Analysis(experiment)

• Performance comparison with various schemes

• Protections with presence of misbehaved hosts

• Bandwidth utilization differentiation• Setting of marking probabilities• A case study of multiple types of traffic

Analysis(experiment)

• Performance comparison with various schemes

• Protections with presence of misbehaved hosts

• Bandwidth utilization differentiation• Setting of marking probabilities• A case study of multiple types of traffic

Performance comparison with various schemes

• to investigate the effectiveness and system stability of our congestion control

Analysis(experiment)

• Performance comparison with various schemes

• Protections with presence of misbehaved hosts

• Bandwidth utilization differentiation• Setting of marking probabilities• A case study of multiple types of traffic

Protections with presence of misbehaved hosts

• we consider the situation where LAN sources do not regulate their rates based on the notification instructions

Analysis(experiment)

• Performance comparison with various schemes

• Protections with presence of misbehaved hosts

• Bandwidth utilization differentiation• Setting of marking probabilities• A case study of multiple types of traffic

• we analyze the bandwidth differentiation feature of our proposed system

Bandwidth utilization differentiation

Bandwidth utilization differentiation

Analysis(experiment)

• Performance comparison with various schemes

• Protections with presence of misbehaved hosts

• Bandwidth utilization differentiation• Setting of marking probabilities• A case study of multiple types of traffic

Setting of marking probabilities

• Although setting a higher marking probability can ensure a stable buffer level around equilibrium threshold as well, this is unnecessary due to the more overhead it may introduce

Setting of marking probabilities

Analysis(experiment)

• Performance comparison with various schemes

• Protections with presence of misbehaved hosts

• Bandwidth utilization differentiation• Setting of marking probabilities• A case study of multiple types of traffic

A case study of multiple types of traffic

• Our scenario deals with uploading of files from clients and a total of three types of traffic are involved

A case study of multiple types of traffic

A case study of multiple types of traffic

1.Client1 and Client2 first connect to ApplicationServer1 sending upload requests of LAN traffic type, and this traffic is light.2. On reception of any request, Application

Server1 connects to Metadata Server for data index

3.&3.5 MetadataServer replies with index to identify and track data storage locations. Traffic load between ApplicationServer1 and MetadataServer can be high

4. After index being obtained, Client1 and Client2 connect to the storage directly for file writing, which is SAN traffic

# HPCServer1 communicates with HPC Server2 with Inter Process Computing (IPC) traffic

A case study of multiple types of traffic

A case study of multiple types of traffic

• Lowest priority : LAN• Higher priority : SAN• Highest priority : IPC

• We adopt a target bandwidth utilization ratio of 1:1.4:1.8 for W1:W2:W3 where W1:W2:W3 describe bandwidth

utilization for LAN, SAN and IPC traffic.

A case study of multiple types of traffic

outline

• Introduction• Ethernet Congestion Management• Differentiated Congestion Control• Analysis (experiment)• conclusion

conclusion

• Our proposed method considered using different AIMD settings for the rate limiter operations and took protection of high priority traffic into account to regulate malicious parties