DC
2
-MTCP: Light-weight Coding for Efficient Multi-path
Transmission in Data Center Network
Jiyan Sun
1
, Yan Zhang
1,*
, Xin Wang
2
, Shihan Xiao
2
, Zhen Xu
1
,
Hongjing Wu
1
, Xin Chen
1
, Yanni Han
1
1
Institute of Information Engineering, Chinese Academy of Sciences, China
2
Stony Brook University, USA
*
Corresponding author, email: zhangyan80@iie.ac.cn
ABSTRACT
Multi-path TCP has recently shown great potential to take
advantage of the rich path diversity in data center networks
(DCN) to increase transmission throughput. However, the
small flows, which take a large fraction of data center traf-
fic, will easily get a timeout when split onto multiple path-
s. Moreover, the dynamic congestions and node failures in
DCN will exacerbate the reorder problem of parallel multi-
path transmissions for large flows. In this paper, we propose
DC
2
-MTCP (Data Center Coded Multi-path TCP), which
employs a fast and light-weight coding method to address
the above challenges while maintaining the benefit of paral-
lel multi-path transmissions. To meet the high flow perfor-
mance in DCN, we insert a very low ratio of coded packets
with a careful selection of the packets to be coded. We fur-
ther present a progressive decoding algorithm to decode the
packets online with a low time complexity. Extensive ns2-
based simulations show that with two orders of magnitude
lower coding delay, DC
2
-MTCP can reduce on average 40%
flow completion time for small flows and increase 30% flow
throughput for large flows compared to the peer schemes in
varying network conditions.
Categories and Subject Descriptors
C.2.1 [
Computer-Communication Networks
]: Net-
work Architecture and Design
Keywords
Data centers; MPTCP; Network coding
1. INTRODUCTION
Today’s data center networks (DCN) provide large-
scale symmetric paths to guarantee high aggregate net-
work bandwidth and transmission reliability [11]. To
fully utilize the bandwidth in DCN, Multi-path TCP
(MPTCP) splits one application flow into multiple sub-
flows and transmits them in parallel on different path-
s [5, 10,21]. Since the link congestions happen randomly
over space and time in DCN [3, 15], the subflows in one
MPTCP connection will experience asynchronous packet
losses (termed APL) [12,21], which results in important
performance issues when applying MPTCP in DCNs.
First, the APL in DCN will seriously reduce the delay
performance of small flows, which take a large portion
of DCN traffic [2, 3, 15, 21, 27]. The small flows often
suffer from high loss probability when large flows occu-
py the link buffer [2]. Specifically, when a small flow
is split into multiple subflows, the average congestion
window of each subflow becomes even smaller, and thus
causing a timeout more easily if any loss occurs [27].
When a timeout happens on a congested subflow, even
though other subflows complete their transmissions first,
the receiver has to wait for 200ms until the lost packet
is retransmitted successfully by its original subflow [6].
Since small flows are often delay-sensitive and their la-
tency are highly related to business profit (e.g., Amazon
estimates every 100ms of latency costs them one percent
profit [19]), such a long delay is generally unacceptable
in DCN [2, 3,15,28].
Second, APL may block the orderly transmissions of
large flows. As the receiver must deliver data to the
application level in sequence, the asynchronous losses
will result in a serious reorder problem. Specially, when
a receiver’s buffer is filled up with the disordered pack-
ets, all the newly received subflow packets have to be
dropped, which therefore blocks the transmissions of
large flows [6, 28].
Network coding has been exploited to alleviate the
APL issues by improving the transmission reliability
through encoding and recovering data packets across
subflows [6,13, 17, 18]. Despite the potential, there are
several important challenges to apply the coding-based
MPTCP designed for general networks [6,17, 18] to the
special environment of DCN. First, the introduced cod-
ing overhead (i.e., the coding delay and the number of
coded packets to insert) should be low enough to meet
the high transmission performance requirements in DCN.
Second, with such a low coding overhead, it poses a great
challenge to make the coded packets effective in improv-
ing the flow throughput and reducing the transmission
time in the presence of dynamic DCN traffic.
To address the above challenges, in this paper, we
propose DC
2
-MTCP, a multi-path transmission protocol
with a light-weight coding technique for subflows to co-
ordinately recover the loss, which significantly improves
the performance for both delay-sensitive small flows and
1