Tcp variants for data center networks

TCP VARIANTS IN DATA CENTER
NETWRKS(ANY 4)
DEPARTMENT OF INFORMATION
TECHNOLOGY
NATIONAL INSTITUTE OF TECHNOLOGY
Under the Supervision of Tulsi sahu
Mr. S. Majumdar Roll No. 16265020
Assistant Professor M.TECH(3rd sem)
Dept. of Electronics and Telecommunication 1

INDEX
 INTRODUCTION
 CHALLANGES IN BRIEF FOR TCP IN DATA
CENTER NETWORKS
1.TCP INCAST
2.TCP OUTCAST
3.QUEUE BUILDUP
4.BUFFER PRESSURE
 TCP VARIANTS FOR DATA CENTER NETWORKS
1. FINE GRAINED TCP RTO
2. FINE GRAINED TCP RTO + DELAYED ACKs DISABLED
3.DCTCP : DATA CENTER TCP
4.ICTCP: INCAST CONGESTION CONTROL FOR TCP
 CONCLUSION

INTRODUCTION
 Transmission control protocol is one of the
most dominant transport protocols and is
widely used by a large variety of internet
application and hence , constitutes majority of
the traffic in data center networks.

THE TRAFFIC IN THE DATA CENTER NETWORKS CAN BE
CLASSIFIED INTO THREE TYPES-

 Thus, bursty query traffic, delay sensitive cat
traffic and throughput elephant traffic co-exist
in data center network.
 Due to these traffic Data center network
require high burst tolerance, low latency and
high throughput.

 The TCP fails to satisfy these requirements
together within the time boundaries because of
impairments such as -
 TCP INCAST
 TCP OUTCAST
 QUEUE BUILD-UP
 BUFFER PRESSURE

CHALLANGES FOR TCP IN DAA CENTER
NETWORK
 The four problems described above are the major
challenges faced by TCP in Data center network-
 1) TCP INCAST-
Results--
1.Packet loss
2.Fast retransmit mechanism is not possible

2.TCP OUTCAST-
Results-
1.The usage of drop-tail queues in switches and
2. Many-to-one communication pattern which leads to
a large set of flows and a small set of flows arriving
at two different input ports ancompeting for the same
bottleneck output port.

3. QUEUE BUILDUP-
This happens when mice traffic and elephant traffic
traverse through the same route.
following are two ways in which the performance of mice
traffic is degraded due to the presence of elephant traffic –
(i)since most of the buffer is occupied by elephant traffic,
there is a high probability that the packets of mice
traffic get
dropped. the implications of this situation are similar to
that of tcp incast because the performance of mice
traffic is largely affected by frequent packet losses and
hence, the timeouts.

(ii) the packets of mice traffic, even when none
are lost, suffer from increased queuing delay
as they are in queue behind the packets of
elephant traffic. This problem has been termed
as Queue build-up.

4.Buffer pressure-
Buffer pressure is yet another impairment
caused by the long lasting and greedy nature
of elephant traffic.
When both mice trafffic and elephant traffic co-
exist on the same route,most of the buffer
space is occupied by packets from the
elephant traffic. This leaves a very little room to
accommodate the burst of mice traffic packets
arising out of many-to-one communication
pattern.
.

The result is that large number of packets from
mice traffic are lost, leading to poor
performance.
Moreover, majority of the traffic in Data Center
Networks is bursty and hence, packets of
mice traffic get dropped frequently because
the elephant traffic lasts for a longer time and
keeps most of the buffer space occupied.

TCP VARIANTS-
 Recently, a few TCP variants have been
proposed for data center networks.
 The major goal of these TCP variants is to
overcome the above mentioned impairments
and improve the performance of TCP in data
center networks.

TCP VARIANTS FOR DATA
CENTER NETWORKS
a) Fine grained TCP RTO
1. The default value of minimum RTO(Retransmit Time
Out) in TCP is generally 200ms. However, it is
significantly larger than the average RTT in a data center
which is in the order of a few micro-seconds.

 sender reciever
packet
ack
200ms
(RTO)
packet
packet
?

 Large number of packet losses due to TCP Incast,
TCP Outcast, queue build-up and buffer pressure
results in frequent timeouts and in turn , lead to
missed deadlines and significant degradation in
the performance of TCP.
 It has been proven that reducing the minimum
RTO from 200ms to 200 𝜇s significantly alleviates
the problems of TCP in simulations and improves
the overall throughput by several orders of
magnitude.

Advantages: The major advantage of this
approach is that it requires minimum
modification to the traditional TCP and hence,
can be easily deployed without any further
complexity.

b)Fine grained TCP RTO + Delayed ACKs
disabled
1. When delayed ACKs are enabled, the receiver
sends only one ACK for every two data packets
received. If only one packet is received, the
receiver waits for delayed ACK timeout period
before sending an ACK.
2. This timeout period is usually 40ms. This
scenario may lead to false retransmissions if fine
grained RTO timers (as explained in the previous
section) are deployed.

3.The reason is that receiver waits for 40ms before
sending an ACK for the received packet and by
that time, fine grained RTO which is in order of
few microseconds, expires and forces the sender
to retransmit the packet.
4.Thus, either the delayed ACK timeout period must
be reduced to a few microseconds or must be
completely disabled while using fine grained
RTOs to avoid such spurious retransmissions.

Tcp variants for data center networks

c) ICTCP: Incast Congestion Control
for TCP
1. The main idea of ICTCP is to avoid packet
losses due to congestion rather than recovering
from the packet losses.
2. It is well known that a TCP sender can send a
minimum of receiver window (rwnd) and
congestion window (cwnd) i.e. min(rwnd, cwnd)).
3. ICTCP influences this property and efficiently
varies the rwnd to avoid TCP Incast.

d) IA-TCP: INCAST AVOIDANCE
ALGORITHM FOR TCP -
1.Unlike ICTCP which use window based
congestion control, IA-TCP uses rate based
congestion control algorithm to control the total
number of packets injected in the network.
2.The motivation behind selecting rate based
congestion control mechanism is that window
based congestion control mechanisms in Data
Center Networks have limitations in terms of
scalability i.e., number of senders in parallel.

3.The main idea of IA-TCP is to limit the total
number of outstanding data packets in the
network so that it does not exceed the
bandwidth-delay product (BDP).
4. IA-TCP adds delay, Δ, to the ACK packet to
ensure that the aggregate data rate does not
exceed the link capacity. Moreover, IA-TCP
also uses delay, Δ, to avoid the
synchronization among the worker nodes
while sending the data.

CONCLUSION
1.TCP, which has been a mature transport protocol
of Internet since past several decades, suffers
from performance impairments such as TCP
Incast, TCP Outcast, Queue build-up and Buffer
pressure in Data Center Networks.
2.we have described each of the above mentioned
impairment in brief along with the causes and
possible approaches to mitigate them. Moreover,
we have carried out a comparative study of TCP
variants which have been specifically designed for
Data Center Networks.

References-
[1]Rohit P. Tahiliani, Mohit P. Tahiliani and K. Chandra Sekaran TCP Variants for
Data Center Networks: A Comparative Study
[2] M. Alizadeh, A. Greenberg, D. A. Maltz, J. Padhye, P. Patel,
B. Prabhakar, S. Sengupta, and M. Sridharan, “Data Center TCP (DCTCP) ,”
SIGCOMM Computer Communications Review,vol. 40, no. 4, pp. 63 74, Aug.
2010. [Online]. Available: http://doi.acm.org/10.1145/1851275.1851192
[3] V. Vasudevan, A. Phanishayee, H. Shah, E. Krevat, D. G. Andersen, G. R.
Ganger, G. A. Gibson, and B. Mueller, “Safe and effective Fine-grained TCP
Retransmissions for Datacenter Communication,” SIGCOMM Computer
Communications Review, vol. 39, no. 4, pp. 303–314, Aug. 2009. [Online].
Available:http://doi.acm.org/10.1145/1594977.1592604
[4]P. Prakash, A. Dixit, Y. C. Hu, and R. Kompella, “The TCPOutcast Problem: Exposing
Unfairness in Data Center Networks,” in Proceedings of the 9th USENIX Conference on
NetworkeSystems Design and Implementation, ser. NSDI’12. Berkeley, CA, USA:
USENIX Association, 2012, pp. 30–30. [Online]. Available:
http://dl.acm.org/citation.cfm?id=2228298.2228339

[5]Y. Chen, R. Griffith, J. Liu, R. H. Katz, and A. D. Joseph, “Understanding
TCP Incast Throughput Collapse in Datacenter Networks,” in Proceedings
of the 1st ACM workshop on Research on Enterprise Networking, ser.
WREN ’09. New York, NY, USA: ACM, 2009, pp. 73–82. [Online].
Available: http://doi.acm.org/10.1145/1592681.1592693
[6] H. Wu, Z. Feng, C. Guo, and Y. Zhang, “ICTCP: Incast Congestion
Control for TCP in Data Center Networks,” in Proceedings of the 6th
International Conference, ser. Co-NEXT ’10. New York, NY, USA: ACM,
2010, pp. 13:1–13:12. [Online]. Available:
http://doi.acm.org/10.1145/1921168.1921186
[7] Y. Ren, Y. Zhao, P. Liu, K. Dou, and J. Li, “A survey on TCP Incast in
Data Center Networks,” International Journal of Communication Systems,
pp. n/a–n/a, 2012. [Online]. Available: http://dx.doi.org/10.1002/dac.2402
[8]J. Hwang, J. Yoo, and N. Choi, “IA-TCP: A Rate Based Incast-
Avoidance Algorithm for TCP in Data Center Networks,” ICC 2012,
2012.

Tcp variants for data center networks

More Related Content

What's hot (19)

Similar to Tcp variants for data center networks (20)

Recently uploaded (20)

Tcp variants for data center networks