tdctcp - csie.org

Report
B99705021 李奕德
 Abstract
 Intro
 ECN
in DCTCP
 TDCTCP
 Performance evaluation
 conclusion
 TCP
does not fit in datacenter
 DCTCP may lower throughput due to small
buffer size
 Improved version of DCTCP (called TDCTCP)
 Compare to existed algorithm (DCTCP,
TCPNewReno)
 TDCTCP may have slightly higher delay but
have much better throughput in general
 Abstract
 Intro
 ECN
in DCTCP
 TDCTCP
 Performance evaluation
 conclusion
 Data
1.
2.

flow in datacenter:
large data flow
require high throughput
small data flow
require low latency
Incast problem
 TCP:
provide : reliable, ordered byte stream
does not provide: high throughput with
“simultaneously low delay”
 Other similar solutions:
 DCTCP, TCPNewReno, TDCTCP
 Tested
under OMNeT++ simulator
 Abstract
 Intro
 ECN
in DCTCP
 TDCTCP
 Performance evaluation
 conclusion
 ECN
mechanism
 2-bit used to describe current situation
00: Non ECN-Capable Transport — Non-ECT
10: ECN Capable Transport — ECT(0)
01: ECN Capable Transport — ECT(1)
11: Congestion Encountered — CE
 Abstract
 Intro
 ECN
in DCTCP
 TDCTCP
 Performance evaluation
 conclusion
 A.
Modification of Congestion Avoidance
 B. Resetting α after Delayed ACK timeout
 C. Dynamic Delayed ACK timeout calculation
 Modification
of Congestion Avoidance
 α = fraction of marked packets in one
congestion window
 Indicates current congestion level
= Maximum Segment Size
 Indicates the size of data that can be sent
 MSS
α after Delayed ACK timeout
 delayed ACK timeout: use in TCP to
reduce ACKs send to the sender
 When ACK timeout occur:
1. α is not updated
2. Old α remain high and block increment
of window size
 α is reset to 0 after every delayed ACK
timeout
 Resetting
Set α to 0 when this happen
 Dynamic
Delayed ACK timeout calculation
 DCTCP: small buffer = small congestion
window
 Congestion window reduce to 1, causing ACK
timeouts
 Packet
arrival follows an exponential
distribution
 Packet loss probability in the network is
small
 Use
10 flows to demonstrate
Low variance in window
size
Spend less time in ACK
timeout
 Abstract
 Intro
 ECN
in DCTCP
 TDCTCP
 Performance evaluation
 conclusion
 Throughput
 Fairness
 Delay
 Queue
length
 Variation in Delay
 Variation in Throughput
 Environment
 Throughput
 Better
- single bottleneck, 1Gbps
performance than DCTCP in general
 Throughput
 Better
- single bottleneck, 10Gbps
than DCTCP under smaller K
 Provide same throughput as TCPNewReno in
early stages
 Throughput
 Better
– multi-bottleneck, 10Gbps
than DCTCP under smaller K
 Provide same throughput as TCPNewReno in
early stages
 Fairness-
 better
single bottleneck, measure in JFI
fairness in every scenario
 Fairness-
multi-bottleneck, measure in JFI
 Delay-
single bottleneck , 10 Gbps
 TCPNewReno
is good except high delay
 Delay-
multi-bottleneck , 10 Gbps
 Queue
length
 TDCTCP
is slightly longer than DCTCP
 Variation
in Delay
 Variation
in throughput
 Abstract
 Intro
 ECN
in DCTCP
 TDCTCP
 Performance evaluation
 conclusion
 Modified
DCTCP => TDCTCP
 15% higher throughput than DCTCP
 improved fairness compare to DCTCP
 provides more stable throughput
 queue length is slightly more than that of
DCTCP at 10Gbps
 delay is slightly higher than that of
DCTCP
Any questions??

similar documents