1 / 43

Anycast by DNS over pure IPv6 network

Anycast by DNS over pure IPv6 network. Minghua Chen & Wei Mao EECS, UC, Berkeley {minghua, maowei}@eecs. Outline. Introduction Anycast Anycast by modifying DNS service over pure IPv6 Network WTRT server selection Model Conclusions Acknowledgement. Introduction.

Télécharger la présentation

Anycast by DNS over pure IPv6 network

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Anycast by DNS over pure IPv6 network Minghua Chen & Wei Mao EECS, UC, Berkeley {minghua, maowei}@eecs

  2. Outline • Introduction • Anycast • Anycast by modifying DNS service over pure IPv6 Network • WTRT server selection Model • Conclusions • Acknowledgement

  3. Introduction • Internet grows dramatically • Exponential growth • In 1999, every 10 seconds, a new pc connects to Internet • Distributed information services face a number of problems of scale • excessive server load • wasted bandwidth • excessive latency

  4. The growth of Internet 460 Million users Source: Cerf, based on www.nw.com, Jan 2000

  5. Server replication & caching • Relatively straightforward method to potentially improve client performance and reduce network load • A key issue in realizing such techniques is how to find the best provider of that service • What is difference between the performance of the best case and the worst case scenario?

  6. Performance variance

  7. What is the relation to our topic? • We believe that anycast service is the right way to do “best” server selection. • Combining anycast service and server replication, the scale problems of distributed information services can be resolved

  8. Outline • Introduction • Anycast • Anycast by modifying DNS service over pure IPv6 Network • WTRT server selection Model • Conclusions • Acknowledgement

  9. Concept Of Anycast • Original definition:a stateless best effort delivery of an anycast datagram to at least one host, and preferably only one host, which serves the anycast address. • Here, anycast is a communication paradigm service, which connect the client to the “nearest” node in a set of nodes that have some same properties

  10. Anycast Illustration Typically, client chooses a nearby server.

  11. Anycast realization • Network-layer anycast • Routing to the nearest server using routing distance metric • Comments • Straightforward idea, hard to implement • Need to modified router • Lack of flexibility in selection criteria

  12. Anycast realization (cont.) • Application layer anycast • Query with a name and a client address. Returns the unicast address of the “best” server • Comments • Does not involve modifying router • Flexible selection criteria • Need periodically collect information

  13. Why we choose application layer implementation? • Implementation Simplicity • Flexible selection criteria • e.g. Server load • More appropriate selection criteria • Network layer: Hop count only is not a good metric for node selection • Application layer: can use other significant information than hop count as metric

  14. Hop count is a poor predictor of performance Hop counts RTT

  15. Problem translation • In application-layer anycast • providing anycast service  selecting the “best” server given necessary information • These information may include • Server load • Latency • Available bandwidth • Client preference

  16. Current server selection schemes • Assign server to a client using round-robin method • Balancing server load • NSCA WWW servers use it • Geographically server assignment • Predicted transfer time (PTT) model or similar model

  17. These schemes are not good enough • Server load • Server load has impact on response latency • Retransmission cost • e.g. Congestion • Traffic segregation according to network topology • Keep traffic “local”– principle of scaling

  18. Outline • Introduction • Anycast • Anycast by modifying DNS service over pure IPv6 Network • WTRT server selection Model • Conclusions • Acknowledgement

  19. Why by modifying DNS • One concern of application-layer anycasting • Require bootstrap mechanism • DNS query is almost an essential step to access Internet services • By quiz • By experience • Is one time query enough for an Internet connection?

  20. The probability of the server’s rank change > 4 is 0.15 Optimal selection changes fairly slowly

  21. Why over pure IPv6 network • We believe that IPv6 will be an important competitor of the next generation IP protocol • IPv6 adopts anycast • IPv6 provides strictly aggregated address space • What will happen in this case?

  22. Outline • Introduction • Anycast • Anycast by modifying DNS service over pure IPv6 Network • WTRT server selection Model • Conclusions • Acknowledgement

  23. Our work: Propose the server selection criteria • Goal is to select the “nearest” server. • Possible distance measures: • lantency, server load, available bandwidth, etc. • propose Weighted Total Response Time (WTRT) • TRT is measured from the time the IP request is sent to the time the whole document is received • The less WTRT, the better • Why? Keen to user’s perspective of QoS

  24. Formulation of WTRT • Express WTRT as the following: • Key elements: • w: segregation weight • Latency: time elapsed from request to start of receiving document • E[packets]: expected total number of packets considering loss and retransmission • P_size: average size of TCP packet • BW: available bandwidth of the link path • K1 and k2: constant coefficients to be determined

  25. Segregation Weight (SW) ?

  26. SW (cont.) • Weight the server within the same subnet less in order to avoid traffic through backbone • Key issue: how to know c/s are in the same subnet???

  27. SW (cont.) • In IPv6 addresses are strictly aggregated  • Do longest prefix matching using c/s IP addresses, similar to the algorithm used in CIDR routing • the longer matched prefix, the more “local” the server is to the client

  28. W4 W1<W2<W3<W4 Details of SW W1 W2 W3

  29. Details of SW (cont.) • Value of SW should be the same for the same level of aggregation • Who determine the SW? • SW could be determined at the authorized DNS server for a particular domain name • SW could be applied by local DNS server on the query results, in order to realize policy flow balance control • Encourage/discourage outgoing flow

  30. Latency • How it affect the TRT • Small latency lead to fast response from server. • When document size is small, this dominates the document retrieve time • What determine latency • Current server load • Current link load and link characteristics

  31. How to estimate latency? • Fei et al proposed a hybrid server push/probe scheme to estimate the latency from a client to a server • Their simulation results show that the scheme works fairly well

  32. Available Bandwidth • How it affect the TRT: • Higher bandwidth leads to faster document transfer • What determines the available bandwidth: • The capacity of the c/s path • Present traffic on the path

  33. Available Bandwidth (cont.) • How to estimate it: • R. L Carter proposed bandwidth probing (BPROBE) algorithm to estimate the total capacity • Also proposed CPROBE algorithm to estimate the current traffic • Their results provide a reasonable ground for other existing researches

  34. Packet Loss & Retransmission • Assume: • K original packets need to be transmitted • p is the steady-state packet loss rate. • Packets lost are retransmitted only once. • Binomial model of the # of retrans. Packets leads to E[packets]=K(1+p). • The total size of the document transmitted becomes E[packets]*P_size. • Average size of TCP packet is 403 bytes.

  35. K1 And K2 • Why K1 and K2? • Linear model • How to determine K1 and K2 • By client preference • small document size  larger ratio of K1 / K2 • large document size  smaller ratio of K1 / K2 • By regression method

  36. Comparison to PTT model • WTRT is derived from PTT but more extensive • WTRT • PTT

  37. PTT WTRT An example Heavy server load

  38. Modified DNS query procedure

  39. Modified DNS query procedure (cont.) • Client IP address and preference are sent between DNS servers • DNS server may return several IP addresses to client for further selection • SW can be determined at authorized DNS server or local DNS server • Periodically, the authorized DNS server gather information from server and probes to update its database

  40. Outline • Introduction • Anycast • Anycast by modifying DNS service over pure IPv6 Network • WTRT server selection Model • Conclusions • Acknowledgement

  41. Conclusions • Server selection + anycast service = solution to the problems of scaling of distributed information services • Application layer anycasting has more advantage over network layer anycasting • In application layer anycasting • providing anycast service  selecting the “best” server given server and link characteristics

  42. Conclusions (cont.) • Anycast by modifying DNS service provide an easy bootstrap mechanism • In IPv6, strictly aggregated address space can facilitate segregating traffic according to network topology • WTRT model is more reasonable and extensive than PTT model

  43. Acknowledgement • Jim Guyton @ Apple Co. for some discussion on traffic segregation • F. Yu and L. Yin for the development of the initial idea Thank you very much for your attention !

More Related