Fast TCP: Late Advancements, Issues and Difficulties.

Uploaded on:
Fast TCP: Late Improvements, Issues and Difficulties Raj Jain Washington College In Holy person Louis Holy person Louis, MO 63131 Microsoft Rapid TCP Workshop Redmond, WA, February 4-5, 2007
Slide 1

Fast TCP: Recent Developments, Issues and Challenges Raj Jain Washington University In Saint Louis Saint Louis, MO 63131 Microsoft High-Speed TCP Workshop Redmond, WA, February 4-5, 2007 These slides are additionally accessible on-line at

Slide 2

Overview Our Congestion Research Then versus Now (1980’s versus 2000’s) High-Speed TCPs Top 10 Requirements for a Good Scheme Two New Problems for Congestion specialists

Slide 3

19.2 kb/s 1 Mb/S File exchange time = 5 mins Time = 7 hours Our Congestion Research 1979-1980: High-Speed Network = 10Mbps Ethernet Implicit Congestion Indication: Drop  Congestion. Drop window to 1. The most effective method to modify windows: AIMD Explicit Congestion Indication: DECBit April 1987: ARPA INENG (IETF)- Bit in the parcel header, Increase/Decrease August 1988: Slow begin paper by V. Jacobson

Slide 5

Increase Policy

Slide 6

Decrease Policy

Slide 8


Slide 9

Where were We Then? 10 Mbps ï‚\' 2 km = 10 ï‚\' 10 6 ï‚\' 2 ï‚\' 10 3 ï‚\' 5 ï‚\' 10 - 6 ï‚\' 2 = 200,000 bits = 25,000 bytes = 17 1500B-bundles. Store and forward deferrals >> engendering postponements  Usual window = 8 How you go from introductory window to 8 has some minor impact How you descend had a noteworthy impact

Slide 10

Where are We Now? 1G is on the tablet/desktop 10G is normal in server farm 100G ï‚\' 40 km Ethernet is being institutionalized in IEEE 802.3 n ï‚\' 10G is utilized as a part of metro systems by means of Link Aggregation 10G ï‚\' 4834 miles across the nation = 10 ï‚\' 4834 ï‚\'1.6 ï‚\' 5 ï‚\' 10 - 6 ï‚\' 2 bits = 773.44 Mb = 96 MB = 198000 512B-portions Which ever way you tally from 1 to 198,000 will be moderate...

Slide 11

LFNs (Elephens) RFC1072 (October 1988): LFNs > 10 5 bits = 12 kB = 8 1500B-parcels TCP needs get window = BDP to keep the funnel full Ideal Sender window size = 2BDP to recuperate from mistakes You have to send no less than 3BDP bytes to get the chance to full speed Bandwidth = Receive window/RTT Default TCP cradle size = 64 kB 64 kB window, 200ms RTT  Max rate = 64kB/200ms = 2.5 Mbps

Slide 12

High-Speed TCPs Core Problem : TCP Reno builds its rate too gradually and diminishes it too quick. Arrangement : Rise quicker and descend slower than Reno

Slide 13

High-Speed TCPs HS-TCP , Sally Floyd, Scalable TCP , Tom Kelly, TCP , Steven Low, , Injong Rhee, TCP (LTCP), Hamilton TCP (HTCP), Westwood , Mario Gerla,… Most of these oblige just send-side alterations to TCP

Slide 14

Congestion in Datacenter Networks Bounded deferral transmission capacity item High-speed: 10 Gbps (now) 100 Gbps (future) Short round-excursion delays 1 Mb to 5 Mb delay-transfer speed item Storage Traffic Þ short access times Þ Low postpone Packet misfortune Þ Long timeouts Þ Not alluring IEEE 802.1au Congestion Notification

Slide 15

Top 10 Requirements for a Good Scheme Fast meeting to steadiness in rates Fast union to decency. Corresponding or Max-min Good for bursty activity Þ Fast meeting Efficient operation: minimize unused limit. Minimize shots of switch Q=0 when sources have movement to send Extremely low (or zero) misfortune Predictable execution: No neighborhood minima Easy to convey Þ Small number of parameters Easy to set Parameters appropriate to an extensive variety of system setups connection velocities, activity sorts, number of sources. Relevant to a mixed bag of switch architectures and queueing/booking controls

Slide 16

Transport for Internet 3.0 Internet 3.0 is the up and coming era of Internet 1.0 = First 20 years = ARPAnet (1969-89) Internet 2.0 = 2 nd 20 years = 1989-2009 NSF GENI/FIND extend How might you outline a vehicle layer today? Window versus Rate Layered versus Cross-Layer AIMD versus Explicit Pacing: Removing Burstiness Ref: Raj Jain, “Internet 3.0: Ten Problems with Current Internet Architecture and Solutions for the Next Generation,”

Slide 17

Summary Time to move from verifiable input, AIMD, and window Handling elephants is less demanding. Mice are testing. The vast majority of the web streams are bursty. Velocity of joining to solidness and decency is essential for bursty movement Time to consider activity administration in the cutting edge Internet.

Slide 18

References J. Jiang and R. Jain, “Forward Explicit Congestion Notification (FECN) for Datacenter Ethernet Networks,” IEEE 802.1au interval Meeting, Monterrey, CA, January 24-26, 2007, Raj Jain, “Internet 3.0: Ten Problems with Current Internet Architecture and Solutions for the Next Generation,” in Proceedings of Military Communications Conference (MILCOM 2006), Washington, DC, October 23-25, 2006, R. Jain, "A Timeout Based Congestion Control Scheme for Window Flow-Controlled Networks," IEEE Journal of Selected Areas in Communications, Vol. SAC-4, No. 7, October 1986, pp. 1162-1167. Reproduced in C. Partridge, Ed., "Innovations in Internetworking," Artech House, Norwood, MA 1988,

Slide 19

References 2 D. Chiu and R. Jain : Analysis of the Increase and Decrease Algorithms for Congestion Avoidance. in Computer Networks, Computer Networks and ISDN Systems, vol. 17, pp. 1–14, 1989, K. Ramakrishnan and R. Jain, "A Binary Feedback Scheme for Congestion Avoidance in Computer Networks with Connectionless Network Layer," Proc. SIGCOMM\'88, August 1988, pp. 303-313, R. Jain, D. Chiu, and W. Hawe, "A Quantitative Measure Of Fairness And Discrimination For Resource Allocation In Shared Computer Systems," DEC Research Report TR-301, September 198

View more...