Multipath TCP for heterogeneous environments

Transcription

1 Multipath TCP for heterogeneous environments Takumi Kurosaka a) and Masaki Bandai b) Dept. of Information and Communication Sciences, Sophia University, 4 7 Yombancho, Chiyoda, Tokyo , Japan a) t_kurosaka@sophia.ac.jp b) bandai@sophia.ac.jp Abstract: Out-of-order packet problem incurs severe degradation of throughput performance of Multipath TCP (MPTCP) with heterogeneous communication links. In this paper, we propose an extended version of MPTCP to solve the out-of-order packet problem 1. The proposed method defines two types of ACK: Regular ACK and Extended ACK. When a receiver receives a data packet, the receiver sends a Regular ACK to the original path through which the data is transmitted, and Extended ACKs to all the other paths. If the Extended ACK arrives earlier than the Regular ACK, the proposed method solves the out-of-order packet problem. We evaluate the performance of the proposed method via NS-3 computer simulations, and show that the proposed method can improve the throughput performance when RTT of paths is heterogeneous. Keywords: TCP, Multipath TCP, mobile communication Classification: Internet References [1] A. Ford, C. Raiciu, M. Handaley, S. Barre, and J. Iyengar, Architechtural guidelines for multipath TCP development, IETF RFC 6182, Mar [2] A. Ford, C. Raiciu, M. Handaley, and O. Bonaventure, TCP extensions for multipath operation with multiple addresses, IETF RFC 6824, Jan [3] R. Stewart, Stream contorol transmission protocol, IETF RFC 4960, Sep [4] Y. Chen, Y. Lim, R. J. Gibbens, E. M. Nahum, R. Khalili, and D. Towsley, A measurement-based study of multipath TCP performance over wirelss networks, Proc. of IMC 13, pp , Oct DOI: / [5] The NS-3 Simulator, [6] MPTCP Code Project, 1 Introduction Recent mobile devices have multiple network interfaces: Wi-Fi and 3G/LTE cellular networks. Multipath TCP (MPTCP) [1, 2] is an extension of TCP and 1 The paper has been presented in part at IEEE CCNC

2 enables mobile devices to use multiple network interfaces simultaneously. Differ from other multipath transport protocols such as Stream Control Transmission Protocol (SCTP) [3], MPTCP is easy to be deployed in practice. Application developers can use traditional APIs of TCP and do not require any modifications to their products. MPTCP has a great potential to improve throughput performance by aggregating bandwidth of multiple communication paths (subflows). However, when subflows have different characteristics such as round trip time (RTT). According to [4], the RTT of Wi-Fi subflow is 25 to 50 milliseconds, and that of AT&T LTE subflow is 40 to 200 milliseconds. In such a situation, data packets from subflows frequently arrive at a receiver in a wrong order, and incur throughput performance degradation of MPTCP. This is out-of-order packet problem. When receiving an out-of-order data packet, a receiver keeps the packet in its receive buffer until an in-order data packet arrives. If a receive buffer is filled up with out-of-order data packets, the receiver advertises its zero-window to the sender. When receiving the ACK with zero-window, the sender is blocked sending the next packet until receiving an ACK with nonzero-window. The more characteristics are different between subflows, the longer the sender is blocked sending packets, which incurs severe throughput performance degradation. If the receiver has infinite receive buffer, out-of-order packet problem does not occur. However, mobile devices have memory limitation and cannot utilize such a large receive buffer. In this paper, we propose an extended version of MPTCP with heterogeneous communication links. The proposed method solves the out-of-order packet problem and improves throughput performance. 2 Proposed method The proposed method defines two types of ACK: Regular ACK and Extended ACK. When a receiver receives an in-order data packet, the receiver sends a Regular ACK to the original subflow through which the data is transmitted, and Extended ACKs to all the other subflows. The format of Extended ACK is the same that of Regular ACK and does not have payload. Extended ACK is used for updating Data Sequence Number (DSN) and advertising window size for a sender. On the other hand, Extended ACK is not used for measuring RTT for calculating retransmission timeout value, and for congestion control of each subflow by the sender. Since the size of Extended ACK is much smaller than that of data packet, the proposed method has a small overhead. When the Extended ACK arrives at the sender earlier than the Regular ACK, the Extended ACK decreases the blocking duration of sending, and makes the sender restart sending next packets more quickly. Algorithm 1 shows the algorithm of the proposed method. The algorithm shows how a receiver makes a decision whether to send Extended ACK when it receives a data packet. k is the index of subflow through which the data packet is sent. N is the number of subflows in a connection. We assume N is a small value such as two. SSN k (Subflow Sequence Number at Subflow k ) and DSN are the values of the sequence number input in the data packet. The receiver keeps 212

3 Algorithm 1 Determination algorithm to send Extended ACK 1: Receive a data packet from Subflow k 2: if SSN k ¼ nextssn k then 3: Calculate new nextssn i 4: lastcount k Count k 5: if DSN ¼ nextdsn then 6: sendflag true 7: if ReceiveBuffer.isEmpty then 8: sendflag false 9: end if 10: Count k 0 11: Calculate new nextdsn and awnd 12: else if DSN > nextdsn then 13: Store the packet to receive buffer 14: Count k Count k þ 1 15: else 16: Count k Count k þ 1 17: end if 18: end if 19: if lastcount k Count k & sendflag then 20: for i ¼ 1 to N do 21: Send an ACK with (nextssn i, nextdsn, awnd) to Subflow i // case i ¼ k, Regular ACK // otherwise, Extended ACK 22: end for 23: else 24: Send a Regular ACK with (nextssn k, nextdsn, awnd) to Subflow k 25: end if nextssn k and nextdsn, respectively. If SSN k ¼ nextssn k and DSN ¼ nextdsn, the data packet is in-order data packet. Otherwise, the receiver realizes that the data packet is out-of-order one. The receiver also keeps a counter, Count k, for counting the number of receiving out-of-order data packet at Subflow k. When the receiver receives an in-order data packet at Subflow k, Count k is reset to zero. When receiving an in-order data packet, the receiver delivers it to application directly and, updates nextssn k and nextdsn. Therefore, if there is a packet whose DSN is equal to the updated nextdsn in a receive buffer, the packet becomes in-order packet and delivered to application directly. In this case, the receiver updates nextdsn again. As a result, the receiver has free space in the receive buffer, and the receiver can advertise nonzero-window size. In this case, Count k is reset to zero, and the receiver sends Regular ACK to Subflow k and Extended ACKs simultaneously to all the other Subflows. If receiving an in-order data packet while the receive buffer is empty, the receiver does not send any Extend ACKs by means of setting sendflag to false. In this case, since the out-of-order packet problem does 213

4 not happen, the receiver does not need to send Extend ACKs. On the other hand, since Extended ACK includes awnd (advertised window size), Extended ACK always advertises nonzero-window size. If there is a Subflow whose RTT is shorter than Subflow k, the Extended ACK sent to the Subflow arrives earlier than Regular ACK. In this case, the sender which is blocked sending next packets can restart sending at the arrival of Extended ACK. If there are some Extended ACKs which arrive earlier than Regular ACK, the proposed method uses only one Extended ACK that arrives at first. In this way, the proposed method decreases the blocking duration of sending, and solves out-of-order packet problem. Fig. 1 illustrates the operation example of the proposed method. In Fig. 1, there are two subflows, Subflow 1 and Subflow 2, between Sender and Receiver. At first, Sender assigns Subflow 2 to transmit a data packet P 1 whose DSN is 1.P 1 is in-order data packet for Receiver. After that, Sender assigns Subflow 1 to transmit three data packets P 2 to P 4 (DSNs are 2 to 4). We assume that the RTT of Subflow 1 (RTT 1 )is shorter than that of Subflow 2 (RTT 2 ). In this case, out-of-order packet problem arises because P 2 to P 4 arrive at Receiver earlier than P 1.AtT zero, we assume that the receive buffer at Receiver is filled up with P 2 to P 4. In this case, Sender is advertised zero-window at T block and blocked sending the next packet. When receiving in-order data packet P 1 at T, the out-of-order situation is resolved. Receiver sends Regular ACK to Subflow 2 and Extended ACK to Subflow 1 simultaneously. Regular ACK arrives at Sender at T 2. Extended ACK arrives at Sender at T 1. In this case, T 1 is earlier than T 2 because we assume RTT 1 < RTT 2. We define the time duration between T block and T 2 as D, and between T block and T 1 as D prop. D also means the blocking duration of sending in conventional MPTCP. D prop is the blocking duration of sending in the proposed method. We define the difference of arrival time between Regular ACK and Extended ACK as D improve. Thus, we can obtain following equation: Fig. 1. Operation example of proposed method. 214

5 D improve ¼ D D prop ¼ T 2 T ðrtt 2 RTT 1 Þ: ð1þ where RTT i is RTT of Subflow i. Since Regular ACK and Extended ACK are sent simultaneously, D improve is almost the half value of the difference between RTT 2 and RTT 1. Eq. (1) implies that the longer difference of RTT between subflows becomes, the more Extended ACK transmitted in the proposed method reduces the blocking duration of sending. In this way, the proposed method effectively improves the throughput performance when RTTs of subflows are different. 3 Performance evaluation We evaluate the performance of the proposed method via computer simulations on NS-3 (version 3.6) [5]. We use [6] as conventional MPTCP. We consider there are two subflows, Subflow 1 and Subflow 2, between a sender and a receiver. We evaluate the throughput performance. Throughput is the aggregated value of all subflows. Simulation results for various RTT 2 s are shown in Fig. 2. The dashed line in Fig. 2 indicates the throughput performance of single TCP using only Subflow 1. We observe that when RTT 1 < RTT 2, the proposed method achieves higher throughput performance than conventional MPTCP. This is because the blocking duration caused by out-of-order packet problem can be effectively reduced by adding Extended ACK in the proposed method. Moreover, when RTT 2 is 250 milliseconds, the proposed method improves the throughput performance up to 58.9%. In this case, the number of Regular ACKs is 22,055, and that of Extended ACKs is 5,203. This means that an overhead of the proposed method is much smaller than the transmitted data size (30 MByte). On the other hand, when RTT 1 ¼ RTT 2, both methods have almost the same throughput performance. This Fig. 2. Throughput comparison result for various RTT 2 s when RTT 1 is 25 milliseconds. 215

6 result leads that the proposed method can obtain almost the same throughput performance as the conventional MPTCP in the ideal condition for MPTCP. 4 Conclusion In this paper, we have proposed an extended version of MPTCP with heterogeneous communication links to solve the out-of-order packet problem. We have confirmed that the proposed method can improve the throughput performance up to 58.9% when RTTs of subflows are different. In addition, we have also confirmed that the proposed method does not affect the throughput performance in the situation when RTTs of subflows are almost the same. As the future work, we modify the proposed algorithm to reduce Extended ACK transmissions when N is large. 216