]> CUHK(SZ)

Shenzhen Guangdong China +86 755 8427 3827 shyang@cuhk.edu.cn

CUHK

Hong Kong Hong Kong SAR China +852 3943 8375 1155136647@link.cuhk.edu.hk

CUHK

Hong Kong Hong Kong SAR China +852 3943 8375 whyeung@ie.cuhk.edu.hk

NCTU

Hsinchu Taiwan China jkzao@ieee.org

General Internet Engineering Task Force BATS code multi-hop This document describes a BATS coding scheme for communication through multi-hop networks. BATS code is a class of efficient linear network coding scheme with a matrix generalization of fountain codes as the outer code, and batch-based linear network coding as the inner code.

This document specifies a BATS code scheme for data delivery in multi-hop networks. The BATS code described here includes an outer code and an inner code. The outer code is a matrix generalization of fountain codes (see also the RapterQ code described in RFC 6330), which inherits the advantages of reliability and efficiency and possesses the extra desirable property of being network coding compatible. The inner code, also called recoding, is formed by linear network coding for combating packet loss, improving the multicast efficiency, etc. A detailed design and analysis of BATS codes are provided in the BATS monograph. The BATS coding scheme can be applied in multi-hop networks formed by wireless communication links, which are inherently unreliable due to interference. Existing transport protocols like TCP use end-to-end retransmission, while network protocols such as IP might enable store-and-forward at the relays, so that packet loss would accumulate along the way. The BATS coding scheme can be used for various data delivery applications like file transmission, video streaming over wireless multi-hop networks, etc. Different from traditional forward error correcting (FEC) schemes that are applied either hop-by-hop or end-to-end, the BATS coding scheme combines the end-to-end coding (the outer code) with certain hop-by-hop coding (the inner code), and hence can potentially achieve better performance. The coding scheme described here can be used in a network with multiple communication flows. For each flow, the source node encodes the data for transmission separately. Inside the network, however, it is possible to mix the packets from different flows for recoding. In this document, we describe a simple case where recoding is performed within each flow. Note that the same encoding/decoding scheme described here can be used with different recoding schemes as long as they follow the principle as we illustrate in this document. The purpose of this document is a preliminary BATS coding scheme for researchers and engineers so that they can develop further the network communication applications/protocols based on BATS codes. Towards a sophisticated BATS code based network protocol, several important research directions should be considered. One is about the security issues. Another important current research is to design the corresponding congestion control and routing algorithms for BATS codes.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

A BATS coding scheme includes an outer code encoder (also called encoder), an inner code encoder (also called recoder) and a decoder. The BATS coding scheme can be used for a single data flow that includes a single source and one or multiple destinations. Thus there exists only one encoder with multiple recoders and decoders. The BATS coding scheme described in this document can be used by a Data Delivery Protocol (DDP) with the following procedures. Outer Code Encoding at a source node which has the data for transmission: The DDP provides the data to be delivered and the related information to the BATS encoder. The BATS encoder generates a sequence of batches, each consisting of a set of coded packets and the information pertaining to the batch. The batches generated at the source node are further recoded before transmitting: A BATS recoder generates recoded packets of a batch. The DDP forms and transmits the DDP packets using the batches and the corresponding batch information. Recoding at an intermediate node that does not need the data: The DDP extracts the batches and the corresponding batch information from its received DDP packets. A BATS recoder generates recoded packets of a batch. The DDP forms and transmits DDP packets using the recoded packets and the corresponding batch information. Decoding at a destination node that needs the data: The DDP extracts the batches and the corresponding batch information from its received DDP packets. A BATS decoder tries to recover the transmitted data using the received batches. The DDP sends the decoded data to the application that needs the data.

Suppose that the DDP has F octets of data for transmission. We describe the procedures of one BATS session for transmitting the F octets. There is a limit on F of a single BATS session. If the total data has more than the limit, the data needs to be transmitted using multiple BATS sessions. The limit on F of a single BATS session depends on the MTU (maximum transmission unit) of the network, which MUST be known by the DDP. We have F is no more than (MTU-10)2^16-1 octets.

The DDP first determines the following parameters: Batch size (M): the number of coded packets in a batch. Recoding field size (q): the number of elements in the finite field for recoding. q is 2 or 2^8 BATS payload size (TO): the number of payload octets in a BATS packet, including the coded data and the coefficient vector. Based on the above parameters, the parameters T, O and K are calculated as follows: O: the number of octets of a coefficient vector, calculated as O = ceil(M*log2(q)/8). T: the number of data octets of a BATS packet, calculated as T = TO - O. K: number of source packets, calculated as K = floor(F/T)+1. The data MUST be padded to have T*K octets, which will be partitioned into K source packets b[0], ..., b[K-1], each of T octets. In our padding scheme, b[0], ..., b[K-2] are filled with data bits, and b[K-1] is filled with the remaining data octets and padding octets. Let P = K*T-F denote the number of padding octets. We use b[K-1, 0], ..., b[K-1, T-P-1] to denote the T-P source octets and b[K-1, T-P], ..., b[K-1, T-1] to denote the P padding octets in b[K-1], respectively. The padding process is shown in .

= T - 1 do bl[Z...T-1] = i break bl[Z...Z+i-1] = i Z = Z + i ]]>

The DDP provides the BATS encoder with the following information: Batch size (M): the number of coded packets in a batch. Recoding field size (q): the number of elements in the finite field for recoding. MAX_DEG: the size of DD. The degree distribution (DD), which is an unsigned integer array of size MAX_DEG+1. A sequence of batch IDs (j, j = 0, 1, ...). Number of source packets (K). Packet size (T): the number of octets in a source packet. The source packets (b[i], i = 0, 1, ..., K-1). Using this information, the (outer code) encoder generates a batch for each batch ID. For the batch ID j, the encoder returns the DDP that contains a sparse degree d[j], and M coded packets (x[j,i], i =0, 1, ..., M-1), each containing TO octets. The DDP will use the batches to form DDP packets to be transmitted to other network nodes towards the destination nodes. The DDP MUST deliver with each coded packet its d: sparse degree BID: batch ID The DDP MUST deliver the following information to each recoder: M: batch size M q: recoding field size The DDP MUST deliver the following information to each decoder: M: batch size q: recoding field size K: the number of source packet T: the number of octets in a source packet The BID is used by both recoders and decoders. The BATS payload size TO MUST be known by all the nodes. The DDP will also include some necessary extra information in the packet header so that the network nodes can identify different BATS sessions, and different end-to-end communication flows. However, such specifications are beyond the scope of this document.

Both the source node and the intermediate nodes perform recoding on the batches before transmission. At the source node, the recoder receives the batches from the outer code encoding procedure. At an intermediate node, the DDP receives the DDP packets from the other network nodes, and should be able to extract coded packets and the corresponding batch information from these packets. The DDP provides the recoder with the following information: the batch size M, the recoding field size q, a number of received coded packets of the same batch, each containing TO octets, and link statistics, e.g., packet loss rates. For a received batch, the recoder determines a positive integer Mr, the number of recoded packets to be transmitted for the batch. The recoder uses the information provided by the DDP to generate Mr recoded packets, each containing TO octets. The DDP uses the Mr recoded packets to form the DDP packets for transmitting.

A destination node needs the data transmitted by the source node. At the destination node, the DDP receives DDP packets from the other network nodes, and should be able to extract coded packets and the corresponding batch information from these packets. The DDP provides the decoder with the following information: M: batch size, q: recoding field size, K: the number of source packets T: the number of octets of a source packet A sequence of batches, each of which is formed by a number of coded packets belonging to the same batch, with their corresponding batch IDs and degrees. The decoder uses this information to decode the K source packets. If successful, the decoder returns the recovered K source packets to the DDP, which will use the K source packets to form the F octets data. The recommended padding process is shown as follows:

The recommendation for the parameters M and q is shown as follows: When q=2, M=16,32,64 When q=256, M=8,16,32,64 It is RECOMMENDED that K is at least 128. However, the encoder/decoder SHALL support an arbitrary positive integer value less than 2^16.

A DDP can form a DDP packet with a header (5 octets), a footer (3 octets) and a payload (TO octets). A DDP packet has totally 8+TO octets.

The BATS packet header has 40 bits (5 octets) and includes fields Packet_Count, Mq, Batch_ID, and Degree.

Packet_Count: 16-bit unsigned integer, specifying the number K of packets of the BATS session. Mq: 4-bit unsigned integer to specify the value of M and q as . Batch_ID: 12-bit unsigned integer, specifying the batch ID BID of the batch the packet belonging to. Degree: 8-bit unsigned integer, specifying the batch degree d of the batch the packet belonging to. MqMqO 00101622 01003224 01106428 000182568 00111625616 01013225632 01116425664

The payload has TO octets, where the first O octets contain the coefficient vector and the remaining T octets contain the coded data. Information in both fields MAY be encoded in JSON (ASCII) or protobuf (binary) formats. coefficient vector: O octets. The range of the value of O is in . coded data: T octets. T is at most MTU - 10, where 10 is the total of the header and footer length plus the minimum value of O.

The footer has three octets. signature: 2 octets. A signature of the individual packet to prevent pollution attack. parity check: 1 octet. A parity check field used to verity the correctness of the packet.

The T octets of a source packets are treated as a column vector of T elements in GF(256). Linear algebra and matrix operations over finite fields are assumed in this section. Suppose that a pseudorandom number generator Rand() which generates an unsigned integer of 32 bits is shared by both encoding and decoding. The pseudorandom generator can be initialized by Rand_Init(S) with seed S. When S is not provided, the pseudorandom generator is initialized arbitrarily. One example of such a pseudorandom generator is defined in RFC 8682. A function called BatchSampler is used in both encoding and decoding. The function takes two integers j and d as input, and generates an array idx of d integers and a d x M matrix G. The function first initializes the pseudorandom generator with j, sample d distinct integers from 0 to K-1 as idx, and sample d*M integers from 0 to 255 as G. See the pseudocode in .

Define a function called DegreeSampler that return an integer d using the degree distribution DD. We expect that the empirical distribution of the returning d converges to DD(d) when d < K. One design of DegreeSampler is illustrated in .

= CDF[d] do return min(d,K) return min(MAX_DEG,K) ]]>

Let b[0], b[1], ..., b[K-1] be the K source packets. A batch with BID j is generated using the following steps. Obtain a degree d by calling DegreeSampler with input j. Obtain idx and G[j] by calling BatchSampler with input j and d. Let B[j] = (b[idx[0]], b[idx[1]], ..., b[idx[d-1]]). Form the batch X[j] = B[j]*G[j], whose dimension is T x M. Form the TO x M matrix Xr[j], where the first O rows of Xr[j] form the M x M identity matrix I with entries in GF(q), and the last T rows of Xr[j] is X[j]. See the pseudocode of the batch generating process in .

The inner code comprises (random) linear network coding applied on the coded packets belonging to the same batch. At a particular network node, recoded packets are generated by (random) linear combinations of the received coded packets of a batch. The recoded packets have the same BID, sparse degree and coded packet length. The number Mr of recoded packets for a batch is decided first by the recoder. Mr can be set as M. When the link statistics is known, the recoder can try to obtain the link packet loss rate e for the link to transmit the recoded batch, and set Mr to be (1+e)M. Suppose that coded packets xr[i], i = 0, 1, ..., r-1, which have the same BID j, have been received at an intermediate node. Using the recommended packet format, it can be verified whether the corresponding packet headers of these coded packets are the same. Then a recoded packet can be generated by one of the following two approaches: forwarding: when receiving xr[i], directly use xr[i] as a recoded packet. linear combination recoding: (randomly) choose a sequence of coefficients c[i], i = 0, 1, ..., r-1 from GF(q). Generate c[0]xr[0]+c[1]xr[1]+...+c[r-1]xr[r-1] as a recoded packet. A recoder can combine these two approaches to generate recoded packets. For example, the recoder will output xr[i], i = 0, 1, ..., r-1 as r systematic recoded packets and generate Mr-r recoded packets using linear combinations of randomly chosen coefficients.

The decoder receives a sequence of batches Yr[j], j = 0, 1, ..., n-1, each of which is a TO-row matrix over GF(256). The degree d[j] of batch j is also known. Let Y[j] be the submatrix of the last T rows of Yr[j]. When q = 256, let H[j] be the first M rows of Yr[j]; when q = 2, let H[j] be the matrix over GF(256) formed by embedding each bit in the first M/8 rows of Yr[j] into GF(256). By calling BatchSampler with input j and d[j], we obtain idx[j] and G[j]. According to the encoding and recoding processes described in and , we have the system of linear equations Y[j] = B[j]G[j]H[j] for each received batch with ID j, where B[j] = (b[idx[j,0]], b[idx[j,1]], ..., b[idx[j,d-1]]) is unknown. We describe a belief propagation (BP) decoder that can efficiently solve the source packets when a sufficient number of batches have been received. A batch j is said to be decodable if rank(G[j]H[j]) = d[j] (i.e., the system of linear equations Y[j] = B[j]G[j]H[j] with B[j] as the variable matrix has a unique solution). The BP decoding algorithm has multiple iterations. Each iteration is formed by the following steps: Decoding step: Find a batches j that is decodable. Solve the corresponding system of linear equations Y[j] = B[j]G[j]H[j] and decode B[j]. Substitution step: Substitute the decoded source packets into undecodable batches. Suppose that a decoded source packet b[k] is used in generating a undecodable Y[j]. The substitution involves 1) removing the entry in idx[j] corresponding to k, 2) removing the row in G[j] corresponding to b[k], and 3) reducing d[j] by 1. The BP decoder repeats the above steps until no batches are decodable during the decoding step.

This memo includes no request to IANA.

Subsuming both Random Linear Network Codes (RLNC) and fountain codes, BATS codes naturally inherit both their desirable capability of offering confidentiality protection as well as their vulnerability towards pollution attacks.

Since the transported messages are linearly combined with random coefficients at each recoding node, it is statistically impossible to recover the individual messages by capturing the coded messages at any one or small number of nodes. As long as the coding matrices of the transported messages cannot be fully recovered, any attempt of decoding any particular symbol of the transported messages is equivalent to random guessing . The threat towards confidentiality, however, also exists in the form of eavesdropping on the initial encoding process, which takes place at the encoding nodes. In these nodes, the transported data are presented in plain text and can be read along their transfer paths. Hence, information isolation between the encoding process and all other user processes running on the node must be assured. In addition, the authenticity and trustworthiness of the encoding, recoding and decoding program running on all the nodes must be attested by a trusted authority. Such a measure is also necessary in countering pollution attacks.

Like all network codes, BATS codes are vulnerable under pollution attacks. In these attacks, one or more compromised coding node(s) can pollute the coded messages by injecting forged messages into the coding network and thus prevent the receivers from recovering the transported data correctly. The research community has long been investigating the use of various signature schemes (including homomorphic signatures) to identify the forged messages and stall the attacks (see , , ). However, these countermeasures are regarded as being too computationally expensive to be employed in broadband communications. Hence, a system-level approach based on Trusted Computing is proposed as a practical alternative to protect BATS codes against pollution attacks. This Trusted Computing based protection consists of the following countermeasures: Attestation and Validation of all BATS encoding, recoding and decoding nodes in the network. Remote attestation and repetitive validation of the identity and capability of these node based on valid public key certificates with proper authorization MUST be a pre-requisite for admitting these nodes to a network and permitting them to remain on that network. Attestation of all encoding, recoding and decoding programs used in the coding nodes. All programs used to perform the BATS encoding, recoding and decoding processes MUST be remotely attested before they are permitted to run on any of the coding nodes. Reloading or alteration of programs MUST NOT be permitted during an encoding session. Programs MUST be attested or validated again when they are executed in new execution environments instantiated even in the same node. Original Authentication of all coded messages using network level security protocols such as IPsec or Peer Authentication over session-based communication using transport level security protocols such as TLS/DTLS MUST be employed in order to provide Message Origin or Communication Peer Authentication to every coded message sent through the coding network.

In addition to the security consideration, there are other issues to be considered towards a fully functionally DDP based on the BATS coding scheme described in this document. Here we discuss the related routing and congestion control considerations, which are research issues that need further elaborations. The general information theoretical guideline is as follows: The outer code of a BATS code can be regarded as a channel code for the channel induced by the inner code, and hence the routing and congestion control algorithms should try to maximize the capacity of the channel induced by the inner code. The BATS coding scheme is flexible for implementing multipath data transmission. Different batches can be transmitted on a different path between a source node and a destination node. For multicast, however, we may allow the combination of certain batches during recoding to improve the multicast throughput. Moreover, to benefit from multiple source nodes, we would need different source node generates statistically independent batches. Congestion control for a DDP employing a BATS code scheme has several extra design considerations. First, the end-to-end throughput should be measured by the average rank rate (the total rank of all the received batches over the time). Second, both the rate of transmitting batches at the source nodes and the number of recoded packets generated by recoding should be adjusted for congestion control. Note that the number of recoded packets can be different for different batch and at different node. Last but not the least, for scenarios like wireless multihop networks, congestion control needs to be performed hop-by-hop, instead of end-to-end as in TCP, due to the high packet loss rate and the long end-to-end delay.

&RFC2119; &RFC8682; &RFC6330;

This becomes an Appendix.