On Pending Interest Table in Named Data Networking

1. On Pending Interest Table in Named Data Networking Huichen Dai, Bin Liu, Yan Chen*, Yi Wang Tsinghua University, China *Northwestern University, USA

2. Outline • Background of Named Data Networking (NDN) • Pending Interest Table (PIT) • Problem Statement • Challenges • Measure PIT Size and Access Frequency • NCE to Accelerate PIT Access and Size Shrinking • Evaluation • Conclusion

3. Background of NDN (1/3) • Newly proposed clean-slate network architecture; • Embraces Internet’s function transition from host-to-host communication to content dissemination; • Routes and forwards packets by content names; • Two kinds of packets: • Interest packet and Data packet. • Request-driven communication model (pull);

4. Background of NDN (2/3) • Content Store (CS): the router’s buffer memory that caches Data packets; • Pending Interest Table (PIT): each entry records the name of a pended Interest and the interfaces from which the Interest comes in. • Forwarding Information Base (FIB): NDN forwarding table.

5. Background of NDN (3/3) Content Store Forwarding Information Base Pending Interest Table Figure: Interest Packet lookup and forwarding process Figure: Data packet lookup and forwarding process

6. Pending Interest Table (1/3) • A special table in NDN and no equivalent in IP. • Keeps track of the Interest packets that are received but yet un-responded. • Brings NDN significant features: • communication without the knowledge of host locations; • loop and packet loss detection; • multipath routing support; etc.

7. Pending Interest Table (2/3) • Read and write on PIT. Interest Packet Data Packet com/google/maps com/google/maps/data/1/ com/google/scholar lookup com/google/scholar/data/1/ PIT Content name Interface list inserted deleted 1 deleted inserted 3 ...... …… A Dynamic Process!

8. Problem Statement (1/2) • A router inserts every incoming Interest into PIT, and removes each received Data packet from PIT. • High-speed packet arrival rate leads to a large PIT size and extremely high access (lookup, insert and delete) frequency. • Two major problems about PIT arise: • The size of PIT? • The access (lookup, insert and delete) frequency of PIT? Size? Frequency?

9. Problem Statement (2/2) • Senior researchers (Lixia Zhang, Van Jacobson, etc.) witness Internet’s function transition; • NDN is empiricallyproposed based on their experience; • A gap from experience to practice: can PIT be satisfied by off-the-shelf technologies? • No previous study touched the feasibility issue. • No statistics support.

10. Challenges • No NDN deployment; • No data; • No PIT;

11. A fresh look at NDN (1/2)-- from application perspective • NDN better satisfies the requirements of users: • browsing web pages, • watching videos, • sharing files, etc. • Users demand the same network applications, but require different implementations in NDN. • The PIT entries are consumed by application-triggered requests.

12. A fresh look on NDN (1/2)-- PIT entry filled by app-triggered Interests • A sample PIT: PIT Content name Interface list com/google/maps 1 HTTP Interest com/google/scholar 3 Avatar 2 P2P Interest parc.com/email_service/van/example@gmail.com/123 5 Email Interest …… …… The applications do not change, solve the problems from the application perspective!

13. Apps from IP to NDN (1/5) • Major Internet applications: • HTTP, FTP, P2P, Email, Online games, Streaming media, Instant messaging, etc. • Implement these applications using NDN’s request-driven communication model. • Afterwards, examine how these applications construct and destruct PIT entries. • HTTP/1.1 as an example.

14. Apps from IP to NDN -- HTTP (2/5) • A request/response protocol; • A client sends a request over a connection to a server, • The server responds with a message. • HTTP/1.1 adopts Persistent Connections: • HTTP requests and responses can be pipelined on a connection. • Simplest Case:

15. Apps from IP to NDN -- HTTP(3/5) • However, the actual case is complicated. • For HTTP Response: • HTTP response divided into multiple Data packets due to the MTU, • At the corresponding PIT entry is created, and at the entry is removed, • Entry Lifetime: .

16. Apps from IP to NDN -- HTTP(4/5) • For HTTP Request: • HTTP requests can be divide into different categories, including GET, HEAD, POST, PUT, DELETE, TRACE; • GET, HEAD – pull data from server; • POST, PUT – push data to server; • DELETE, TRACE – function calls on the server. • NDN can only pull data from server.

17. Apps from IP to NDN -- HTTP(5/5) • GET, HEAD – implemented by Interest packet directly; • Extend the function of Interest packet to accommodate the rest methods; • POST, PUT – add additional data block to the Interest packet that contains the content to be pushed to the server; • DELETE, TRACE – extend Interest packet to enable function calls.

18. Measure PIT Size and Access Freq. (1/3) • So far, we have • So we can: • We get data – we quantify PIT size and access frequency by analyzing the NDN trace. Applications on IP IP trace Applications on NDN NDN trace translate

19. Measure PIT Size and Access Freq. – Metrics (2/3) • Size: the number of PIT entries at each snapshot by examining the lifetime of each PIT entry. • Lookup frequency: the # of Data packets (excluding the last Data of a request/response pair) arrive per second; • Insertfrequency: the # of emerging biflows(new Interests arrive) • Delete frequency: the # of disappearing biflows (last Data packet of a request/response pair arrive).

20. Measure PIT Size and Access Freq. – Results (3/3) • A one-hour IP trace from a 20 Gbps link in China Education and Research Network (CERNET). • PIT size: 1.5 million • Frequencies of the first 10 seconds: (1.4 M/s, 0.9 M/s, 0.9 M/s) • Methodology is what matters.

21. Data Structure for PIT (1/3) • PIT size and access frequency demand a well-organized data structure to implement PIT. • Inspired by the Trie structure in IP lookup, we organize PIT by a Trie as well. A sample IP trie, bit-wise.

22. Data Structure for PIT (2/3) • NDN name is different from IP address; • IP addresses: • Fixed length, short… • NDN names: • hierarchical, composed of component, enables aggregation; • Unbounded # of components, much longer than IP; • E.g., /com/parc/bulletin/breakthroughs.html • Therefore, we adopt Name Prefix Trie, which is of component granularity.

23. Data Structure for PIT (3/3) level-1 level-2 level-3 level-4 level-5 Name Prefix Trie (NPT), each edge stands for a name component. news 4 yahoo 3 uk maps 5 6 2 com maps google 7 8 1 maps google A B cn sina 9 C map baidu An example of PIT D E Name Prefix Tree

24. Name Component Encoding (1/4) • The Name Prefix Trie (NPT) makes PIT well organized, but matching each component is slow: • The length of each component varies; • should be an exact match of string patterns. • We further propose to use positive integers to encode components – Name Component Encoding (NCE). • An integer occupies smaller memory space compared to a component; • Matching integers is much faster.

25. Name Component Encoding (2/4) level-1 level-2 level-3 level-4 level-5 1 4 3 1 1 5 6 2 2 level-1 level-2 level-3 level-4 level-5 1 1 Name Prefix Trie (NPT) is transferred to Encoded Name Prefix Trie (ENPT). 2 7 8 news 4 1 yahoo 3 uk 1 3 maps A B 5 6 2 2 2 9 C com maps google 7 8 1 1 D E 1 maps google A B cn sina 9 C map baidu D E CAS Name Prefix Tree Encode Name Prefix Tree

26. Name Component Encoding (3/4) • Preparation: define the edges in the NPT leaving from the same node as a Code Allocation Set (CAS) (dotted ellipse on the NPT); • The rules to assign codes to components: • Assign each name component in a CAS a unique code; • The code should be as small as possible (ensures the codes are continuous). • The ENPT is a logical structure and is eventually implemented by Simplified State Transition Arrays (S2TA). • Benefit: a component match on S2TA can be implemented by a single memory access with the codes as indexes!

27. Name Component Encoding (4/4) level-1 level-2 level-3 level-4 level-5 1 4 3 1 1 5 6 2 2 1 1 2 7 8 1 1 3 A B 2 2 9 C 1 1 D E Simplified State Transition Array (S2TA) Encode Name Prefix Trie (ENPT) The data structure that really implements the ENPT, each transition stands for a node and its outgoing edges. The codes are indexes of each outgoing edge.

28. Evaluation • Two name sets to compose two PITs: • Domain names collected from ALEXA, DMOZ and our web crawler, around 10 Million; • URLs extracted from the HTTP requests in the trace, around 8 Million. • Evaluation goals: • memory consumption of S2TA, • access frequency S2TA, • Comparison with other methods.

29. Evaluation Results (1/3) • Memory Consumption Figure: 10M Name Set memory consumption–original size, NPT size, ENPT. compression ratio (10M Name Set): 63.66% Figure: 8M Name Set memory consumption–original size, NPT size, ENPT. compression ratio (8M Name Set): 12.56%

30. Evaluation Results (2/3) • Access frequency Figure: Lookup, insert and delete performance for the 10M Name Set Average: 3.27 M/s, 2.93 M/s and 2.69 M/s Figure: Lookup, insert and delete performance for the 8M Name Set Average: 2.51 M/s, 1.81 M/s and 2.18 M/s

31. Evaluation Results (3/3) Figure: PIT access frequency speedup based on S2TA. All these results reveal that PIT can be implemented by off-the-shelf technologies.

32. Conclusion • Measure the size and access frequency of a specific PIT; • Prove the feasibility of PIT with off-the-shelf technologies; • Propose NCE to accelerate PIT access operations while reducing PIT size.

33. Thanks!Questions please ^_^