Download as pdf or txt
Download as pdf or txt
You are on page 1of 2

Paper Review

Safe and Effective Fine-grained TCP Retransmissions for Datacenter Communication

Muhamad Erza Aminanto


Student ID: 20145562
Cryptology and Information Security Laboratory
Korea Advanced Institute of Science and technology
aminanto@kaist.ac.kr

Abstract—This paper review contains review to “Safe and  The server only responding with a small amount
Effective Fine-grained TCP Retransmissions for Datacenter of data for each request.
Communication” from V. Vasudevan. First, I will explain about
summary of this paper. The main problem of this paper is TCP B. Solutions
incast problem. Then the solution proposed by the Authors is They propose a solution to prevent TCP incast problem.
reducing RTO to 1 ms or lower. After that, I have summarize The solution is reducing RTO sharply until approximately
implementation and evaluation of this solution. Then, I will
have same order to network delay/RTT. Based on their
describe two related research. Afterwards, I have expressed my
opinion about this paper, also about its advantages and
simulation and real-world study, they able to show that
disadvantages. reducing RTO until 1 milliseconds was effective to prevent
TCP incast problem for small amount of concurrent senders
Keywords—TCP incast; RTO; RTT; Data Center; (i.e. 8-16 senders).
C. Implementation and Evaluation Summarize
I. SUMMARY
They are using Linux kernel 2.6.28 for implement their
This section contains 3 parts, first is problem domain that solution. This Linux kernel have limitation that the TCP RTO
the writer want to solve, then solutions offered by them and last may reduced only to 5 ms. So, they modified the kernel to use
is their implementations and evaluations to their solutions. microsecond-accurate timers. After this modification, they can
reduce RTO to 1 ms even in 200 microseconds. The results of
A. Problem Domain their implementation is the RTO should be reduced to 1 ms or
This paper want to solve problem about TCP incast. TCP lower to prevent TCP incast problem. Actually, their
incast problem is the throughput was falling down when implementation have some minor differences with their
multiple clients communicate to single server in TCP simulation as shown in Figure 2 and Figure 3. The differences
networks that have high-bandwidth and low delay. The very was caused by some factors, like RTT variances and physical
fast data transmissions overfill Ethernet switch buffers that delay. Although there are minor differences, this research still
causing intense packet loss, then leads to TCP timeouts or able to shown that reduced RTO may prevent TCP incast
Retransmission Time Out (RTO). The main concern of the problem. Also, the Author concerned about the importance of
writer was the value of RTO usually 200 milliseconds while desynchronizing retransmissions by adding some randomness
most of Round-Trip-Time (RTT) value is tens or hundreds
to the RTO. It becomes necessary to avoid throughput dropped
microseconds. They thought that this gap is too large/wide,
and they want to minimize this gap. The writer said that they as the amount of low-latency flows growing highly.
faced 3 challenges to solve this problem: first, they want to II. RELATED WORK
show their solution is practical. Second, they want to show
effectiveness of their solution. Last, they want to show that Related paper to this paper is “Taming TCP Incast
their solution is safe to implement in wide-area. Throughput Collapse in Data Center Networks” from J. Zhang.
This paper presents a simple and effective TCP enhanced
Also, the writer stated preconditions for TCP incast occurs mechanism, called GIP (Guarantee Important Packets), for the
in follow conditions: applications with the TCP incast problem. The main idea is
 The networks have high bandwidth and low making TCP aware of the boundaries of the stripe units, and
latency, also switches with small buffers. reducing the congestion window of each flow at the start of
each stripe unit as well as redundantly transmitting the last
 The clients using parallel barrier-synchronized
packet of each stripe unit. GIP modifies TCP a little at the end
requests which only send another request if all
hosts, thus it can be easily implemented.
responses from earlier request have been
received. The second related paper is “Data Center TCP (DCTCP)”
from M. Alizadeh. This paper also concerned about TCP incast
problem. This paper offer solution with modify TCP protocol if this research is implemented to real Data Center with large
by adding Explicit Congestion Notification (ECN). amount of servers.

III. ADVANTAGES AND DISADVANTAGES REFERENCES


The advantages of solution offered in this paper is the [1] V. Vasudevan, et.al., “Safe and Effective Fine-grained TCP
reduced RTO was proofed safe to use in wide area (WAN Retransmissions for Datacenter Communication,” SIGCOMM’09,
August 17–21, 2009, Barcelona, Spain. Copyright 2009 ACM 978-1-
environment). On the other hand, the disadvantages of the 60558-594-9/09/08.
solution is the reduced RTO is can’t prevent dropping [2] M. Alizadeh, et.al., “Data Center TCP (DCTCP),” SIGCOMM’10,
throughput in large number of servers simulation as shown in August 30–September 3, 2010, New Delhi, India.. Copyright 2010.
Figure 6. [3] J. Zhang, et.al., “Taming TCP Incast Throughput Collapse in Data
Center Networks,” IEEE 2013 978-1-4799-1270-4/13.
IV. MY OPINION [4] ______, “TCP Timeout and Retransmission,” [Online]. Available at:
http://www.pcvr.nl/tcpip/tcp_time.htm#21_2
Overall, I like this paper. The Authors have a good and [5] Brad Hedlund, “TCP Incast and Cloud Application Performance,”
structured writing. Also they have explained kernel [Online]. Available at: http://bradhedlund.com/2011/05/01/tcp-incast-
modification in detail. Then, I think ECN might be another and-cloud-application-performance/
solution for this TCP incast problem. Also, maybe it will better

You might also like