Transport Protocols for Data Center Communication. Evisa Tsolakou Supervisor: Prof. Jörg Ott Advisor: Lect. Pasi Sarolahti

Transcription

1 Transport Protocols for Data Center Communication Evisa Tsolakou Supervisor: Prof. Jörg Ott Advisor: Lect. Pasi Sarolahti

2 Contents Motivation and Objectives Methodology Data Centers and Data Center Networks - Architectures - Traffic Characteristics - Requirements TCP Protocol Queue Management and Explicit Congestion Notification DCTCP Protocol Analysis Conclusions Future Work 2

3 Motivation and Objectives Cloud computing is constantly gaining more ground in the provision of services Data Centers and Data Center Networking are becoming increasingly more important Study of Data Centers and Data Center Networking 3

4 Motivation and Objectives Transmission Control Protocol (TCP) Challenges for TCP in the Data Center Environment TCP Improvements and other Traffic designs Queue Management and Explicit Congestion Notification (ECN) Data Center TCP (DCTCP) Analysis of the Protocol 4

5 Methodology Study of the Data Center environment Architectures Data Center Traffic Characteristics Challenges in the Data Center environment Survey of transmission designs and study of signaling (in the Internet and in the Data Centers) Emulated testbed for the study of DCTCP 5

6 Data Centers and Data Center Networking: Architectures 6

7 Data Centers and Data Center Networking: Architectures Fat Tree, DCell and Bcube architectures 7

8 Data Centers and Data Center Networking: Traffic Characteristics The traffic in Data Centers comprises of two main patterns: Traffic between servers Traffic between applications and servers Traffic comprises of small flows, medium sized flows and large updates Congestion occurs very often Most of the flows last less than 10 secs Losses are greater on the aggregation layer Packets arrive in groups 8

9 Data Centers and Data Center Networking: Requirements High burst tolerance Low latency High throughput 9

10 TCP Protocol Characteristics Timeout and Retransmission Procedure Congestion Control 10

11 TCP Protocol TCP segment, three-wayhandshake, AIMD(Additive Increase Multiplicative Decrease) pattern 11

12 TCP Protocol Challenges in Data Centers TCP Incast (Flows synchronization and switch interface of the aggregator overflown with synchronized flows concurrently) TCP Outcast (Large and small flows arrive at different input ports of a switch but compete for the same output port) Queue Buildup (Packets of small flows build queues after packets of large flows) Buffer Pressure (Reduction of Buffer space because of queue buildup) Improvements for Data Center Networks Transport Designs that tackle the Data Center environment challenges 12

13 Queue Management and Explicit Congestion Notification Queue Management Active Queue Management (AQM): Management of buffers before queues become full Random Early Detection (RED) Other AQM mechanisms (CoDel, PIE) ECN Mechanism used to notify of congestion Network components make use of AQM to set a Congestion Experienced codepoint in the IP header ECN in TCP: ECE and CWR flags in TCP header, negotiated use, reaction of the protocol 13

14 DCTCP Protocol TCP variant specifically designed for Data Centers End hosts receive congestion information from network components Algorithm works in three parts: AQM mechanism based on marking parameter K (configurable) Every packet gets ECE marked if it has a CE endpoint mark Sender maintains an estimated fraction of packets that are marked that is updated once every RTT The protocol reacts to congestion by reducing the congestion window based on the amount of congestion experienced The protocol has been found to outperform TCP in Data Center environment scenarios 14

15 Analysis Use of Network Emulation: Virtual test networks Main tool: Mininet Network emulator used to test Software Defined Networking Applications Python API C code for the emulation Use of Linux Kernel 15

16 Analysis Topologies used in Mininet 16

17 Analysis For the different scenarios different parameters were changed, like the number of hosts, the link bandwidth, the delay of the links To see the difference between TCP, TCP with the use of ECN and DCTCP the graphs depict the CWND change and the queue occupancy 17

18 Analysis Comparison of CWND for 100 hosts and 100MBps bandwidth for TCP, ECN TCP and DCTCP 18

19 Analysis Comparison of queue occupancy for 100 hosts and 100MBps bandwidth for TCP, ECN TCP and DCTCP 19

20 Analysis Difference in queue occupancy in TCP and DCTCP traffic 20

21 Conclusions Behavior of DCTCP in links with different characteristics Verification of DCTCP outperforming TCP Lower queue occupancy main benefit from the use of DCTCP DCTCP tackles the Data Center challenges with success Important role of ECN feedback 21

22 Future work Comparison of DCTCP and TCP in actual testbeds Further study of DCTCP for Internet deployment Further study of ECN 22