We consider a large-scale network with n nodes, in which k packets of information are held (eg, k packets of information have been sensed or collected in some other way). Nodes in the network are very vulnerable because of limited energy supply and poor environment, and it hopes to distribute the acquired information throughout the network. Thus, each of n nodes stores a (possibly encoded) packet, and the original k source packets can then be recovered from any (1 + E)k nodes in a computationally simple manner, and for some small small > 0.
Two distributed algorithms are proposed to solve this problem.
1. Random walks
presents a solution using random walks with traps. To obtain the desired code degree distribution, they adopted the Metropolis algorithm to specify the transition probabilities of the random walk.
2. Fountain code
We developed two distributed algorithms based on simple random walk and Fountain code to solve this problem. Unlike all previously developed schemes, our solution is truly distributed, that is, nodes have no knowledge of n, k or connectivity in the network, and they do not maintain any routing tables except in their own neighbourhoods. In the first algorithm, all sensors have knowledge of n and k. In the second algorithm, each sensor estimates these parameters by random walk propagation.
We analyze the communication/transmission and encoding/decoding complexities of the two algorithms and provide extensive simulation results.
Contributions:
1. Two new algorithms, simple random walk and LT code,
compared with others:
the traditional way of knowing n and k in advance.
These algorithms are simpler, more robust, and have fewer constraints.
2. The transmission, encoding and decoding are analyzed.
3. Experimental results
Probability of degree distribution (you can see from the picture that the degree obtained according to different functions is different)
Data Update
Consider the problem of data update.
Data update is generally a change that is difficult to control in the entire ec process, and generally increases write amplification, IO transmission, etc.
The problem becomes simpler if we assume that storage nodes keep the IDs of accepted packets. We are onlyrandom walk, and check the ID of the upcoming packet. Suppose node u keeps track of all IDs of packets it accepts. If the ID of the upcoming packet is already contained in u's ID list, u accepts the updated message. Otherwise, u will forward the packet with the time-to-live counter incremented. If this counter reaches the threshold, the packet will be dropped. The following steps describe the update scheme:
If each node records the id table of a data packet,
if the updated data packet is in the id table, the node accepts the update.
Otherwise the node forwards the packet and increments the time-to-live count.
If the count reaches the threshold, it is discarded directly.
Process:
1. The si node prepares a group
: new and old data packets, id and counter.
Likewise, si adds an update counter token initialized to 1 for the first updated packet.
Therefore, we assume the following steps happen when token is set to 1. (When not found in the ID table)
si randomly selects a neighbor node u, and transmits this data packet k.
2. Encoding process:
node u checks if the data packet k is updated or the first time. If it is the first packet, it will be accepted, forwarded or discarded as shown in LTCDS-I Algorithm 1. If packetsi is a newer packet, node u checks whether IDsi is already included in its accept list. If yes, then it updates its value yu as follows.
If not, it will add the updated packet and add 1 to the counter.
This packet will be discarded when the counter reaches the threshold. The counter must be large enough to be able to find all the old data and update it.
(Isn't it directly equal to the number of nodes???)
3. Stored procedures
If all nodes have finished updating their value yi. A decode stage can be run to retrieve raw data and update information.
Now, since we only run a simple random walk once per update, if h is the number of nodes whose values are updated, we get the following.
performance evaluation
Two values are measured:
the probability of successful decoding
and the decoding rate: the decoding rate η is the ratio between the number of nodes h queried and the number of sources k
Figure 3:
| Figure few | X | y | those lines |
|---|---|---|---|
| 3 | Decoding rate change (1-2.5) | Probability of successful decoding | Different combinations of node n and original data k (n=100, 200, k=10,20,40) |
| 4 | Decoding rate change (1-2.5) | Probability of successful decoding | Different combinations of node n and raw data k (n=200, 500, 1000) |
| 5 | Participating n (500-5000) | Probability of successful decoding | Variation of decoding rate (1.4, 1.7) |
| 6 | Change of system parameter c (0-5, step size 0.5) | probability of successful decoding | Different combinations of node n and original data k (n=500, k=50; n=1000, k=100) |
| 7 | Decoding rate (1-2.5) | probability of successful decoding | different n and k for different methods |
| 8 | Decoding rate (1-2.5) | probability of successful decoding | Different methods for different n and k M |
| 11 | System parameter c (0-60, step size 10) | probability of successful decoding | Different n and different k, the code rate is different |
Conclusions:
These algorithms are simpler, more robust, and have fewer constraints than previous solutions, which required knowledge of the network topology, the maximum degree of nodes, or knowledge of the values of n and k [4, 6, 9 , 10, 11]. We calculated the computational encoding and decoding complexities of these algorithms and simulated their performance with small numbers of k and n nodes. We show that the number of sources and the total number of nodes can be successfully estimated if a node only counts the access time and the time between packets.
Our future work will include distributed network storage algorithms for sensor networks based on Raptor code. We also plan to provide theoretical results and proofs for the results shown in this paper, where limited space is not an issue. Our n and k value estimation algorithm is promising, and we plan to investigate other network models where this algorithm is beneficial and can be exploited.