TCP Congestion Control Evolution

Transcription

1 TCP Congestion Control Evolution (from Steven Low presentation) Junho Suh MMLAB 1

2 Overview Internet History TCP basics & history Some trends in the Internet Implication from some trends MMLAB 2

3 Internet History First 25 years of Internet Networks are several orders of magnitude smaller in size, speed, and heterogeneity Applications are much simpler Last 14 years of Internet Networks are much bigger, faster, and heterogeneous Applications are much more diverse and demanding TCP congestion control gradually becomes unsuitable for more and more applications in more and more networks MMLAB 3

4 Internet History Both networks and applications are evolving much faster in last 14 years than previous 25 Our control mechanisms have shown remarkable robustness, but also essentially frozen in the last 20 years TCP congestion control is gradually being stressed more and more Time to change is now MMLAB 4

5 Birth of Reliable Data Tx TCP Cerf and Kahn, Trans. Communications (1974) RFC 793, TCP (1981) ARPANet cutover to TCP/IP (Jan 1, 1983) Window control: ack-clocking, timeout, retransmission For correct delivery of packet sequence over unreliable networks reliable data transmission To prevent overwhelming receiver (through Advertised Window in TCP header) flow control We envision [HOST retransmission capability] will occasionally be invoked to allow HOST accommodation to infrequent overdemands for limited buffer resources, and otherwise not used much. -- Cerf and Kahn, 1974 MMLAB 5

6 TCP basics Three important mechanisms Congestion control Adjust amount of packets ongoing on networks to avoid network congestion Packet loss indicates network congestion (important assumption) A loss -> duplicate ACK or timeout Loss recovery Reliable data transmission Retransmission Flow control Adjust amount of packets to avoid overflow of receiver s buffer Ordered delivery Send packets in ascending order MMLAB 6

7 TCP Tahoe and its variants TCP Tahoe TCP Reno MMLAB 7

8 TCP Problems The original TCP does not much perform well in high speed TCP and long distance networks Some congestion control protocols have been proposed to remedy The problems are listed below: Scalability RTT fairness TCP friendliness Fairness and convergence time New applications demand new requirements that stretch TCP FTP or Telnet vs video streaming New wireless challenges MMLAB 8

9 Trend 1: Online content is exploding Exploding need to deliver large content Video, software, games, business info For example CAGR : 36% Google worldwide: 46PB/month Library of Congress: 0.136PB MMLAB 9

10 Video Streaming Over Wireless: Where TCP is Not Enough Limitation of TCP for supporting video streaming over wireless, in the presence of heterogeneous link speeds The setup mimics a wireless home network with several nodes operating in ad-hoc mode. We demonstrate the impact of a file transfer session over a slow link on an ongoing video streaming session over a fast link TCP leads to approximately equal throughput for both sessions, despite difference in their link speed, causing severe quality degradation of the video stream. MMLAB 10

11 Why is TCP not good enough for Mobile Operators? In mobile wireless situation TCP suffers substantially when there is a high degree of rate/load asymmetry in the forward and reverse directions TCP is vulnerable when a mobile host has payload traffic in both directions Fairness notion adopted at the wireless schedulers are not necessarily compatible with the per flow fairness of TCP MMLAB 11

12 Trend 2: Broadband penetration Infrastructure growth to support delivery need Global broadband penetration is accelerating 11 million new subscribers/month globally 83% of US home Internet access is broadband by June 2007 MMLAB Source: Deutsche Bank, 12 Jan 2007

13 TCP Challenges in Multi-hop Wireless Networks TCP has not been designed to cope with the complex interference found in static multi-hop wireless (mesh) networks Interference makes congestion and rate control a neighborhood, not a single-node affair MMLAB 13

14 Trend 3: Centralization Centralize IT to reduce costs of management, space, power, cooling Exacerbated by virtualization of infrastructure and personalization of content Power & cooling cost is escalating, fueling centralization CAGR ( ): power & cooling 11.2%, new server spend 2.7% In 2005, 1,000 servers cost $3.8M to power & cool in 4yrs; 2% increase in electricity cost raises cost by $200K MMLAB 14

15 Trend 3: Centralization MMLAB 15

16 TCP Issues in the Data Center The number and size of datacenters are growing exponentially, driven by businesses such as Google, Amazon, and the Wall Street trading firms Essentially all the network applications in these datacenters depend on TCP Yet in datacenters, the bandwidth, latency, cost, and congestion parameters that affect TCP are drastically different than in the wide-area Internet MMLAB 16

17 Conclusion Some Implication from Trends Unless TCP explicitly accounts for this, its performance can be spectacularly bad More large contents over longer distance Served from centralized data centers Need to consider clean state design or evolution You could only get that sustained rate if you are delivering within 100 miles, due to the way current Internet protocols work. Tom Leighton, MIT/Akamai, Oct 2007 MMLAB 17