What is the business quality of CDN & DNS? How to recoginize the traps when developing CDN & DNS system? How to evalute the metrics of business scenarios of CDN & DNS? This slides provide the best practices of developing reagarding CDN & DNS by SpeedyCloud. For more information, please refer to www.speedycloud.cn. You can also mail to : social@speedycloud.cn

1. How to Design and
Develop DNS System in
CDN

2. Agenda
 Intelligent DNS Resolution
 DNS Resolution
 Performance Requirement
 Relationship between Intelligent DNS and CDN traffic scheduling

3. Intelligent DNS Resolution
 Process
 CDN & DNS support standard protocols
 Communication of Authoritative DNS
 System Consuming Characteristics

4. Process

5. CDN & DNS Supports Standard Protocols
 rfc 1035 : basic
 rfc 2671 : EDNS0
 rfc 3596 : AAAA
 ECS supports: draft-vandergaast-edns-client-subnet-04

6. Authoritative DNS Communication
 Typical Messages
 Mainly UDP
 Single package request response
 Small package
 No-repeat five-tuple
 Business aspect
 Delay-sensitive
 Distribution deployment

7. System Consuming
 Network I/O Intensive
 Small package, high frequency
 One-time network I/O
 CPU Intensive
 Small package, high frequency
 One-time network I/O
 Domain mainly for string query and processing
 Extra CDN strategy computing

8. Domain Resolution
 Terminal interaction with local DNS
 Local DNS interaction with CDN DNS
 Local DNS interaction with DNS cluster

9. Terminal interaction with local DNS
 Most terminal uses local DNS
Local terminal users
Local DNS terminal users

10. Local DNS interaction with CDN DNS
 The user scale of local DNS varies significantly.
 CDN DNS perspective
 Local DNS cache dilutes hotspots
 Traffic scheduling perspective
 Each DNS visits cover varying sizes of users
CDN DNS
LDNS用户群

11. Local DNS interaction with DNS cluster
 Local DNS Picking：Resolving Delay
Sensitive
 Choose near NS for high probability
(short RTT)
 Choose far NS for low probability (long
RTT)
 Traffic scheduling perspective
 Each NS has unequal visits

12. Local DNS
Example: How to choose the best NS
0
5
10
15
20
25
30
35
40
45
50
DNSDevices
Delayms
DNS Resolution Sequences

13. Local DNS
Example: Punishiment under high delay
502502
0
50
100
150
200
500 510 520 530 540 550 560 570 580 590 600 610 620 630 640 650 660 670 680 690 700
DNSDevices
Delayms
DNS Resolution Sequences

14. Performance Requirement
 High quality DNS system’s performance
 Performance index evaluation
 Performance test
 Network I/O selection
 New DNS data storage

15. High quality DNS system’s performance
 Resemble ICMP echo services (ping)

16. Performance index evaluation
Performance
C40M - 40Gb
C10M - 10Gb
C1M - 1Gb
C500K-C100 - 100Mb
C10K - 10Mb
C1K - 1Mb
C100K
Domain Resolution
Quanlity
Huge DDoS
Small DDoS
DDoS
Huge website peak analysis
Active website
Single Machine
Performance
Knot、NSD、BIND、PowerDNS
BIND+DLZ+NOSQL

17. Hot-spot dilution’s impact on performances
 Reason of hot-spot dilution
 Upstream Local DNS is cache-like services, which won’t be visited in before
expiration
 Hot-spot dilution effects
 Online performance much lower than test performance
 High cache missing of software and hardware
 In Cloud, Higher cache missing of software and hardware
 Beware
 Extra resources need to be spared during evaluation
 On storage and algorithm selection, extra resources need to be spared for
random access performance stability
 Replay or clone online traffic for DNS test.

18. Performance Test
 queryperf
 DNSPerf
 tcpreplay
 tcpcopy
 Performance test devices

19. Traps in performace evaluation
 What is the parallel performance of CDN DNS?
 Authoritative DNS’s purpose is to finish one-time response ASAP
 For CPU intensive business, better lower the parallel operation to avoid getting out of
control
 QPS is the principal index on evaluating CDN DNS’s performance

20.  How about using queryperf/DNSPerf to evaluate online CDN DNS performance?
 Hard to cover the Local DNS variation of CDN DNS
 Fixed five-tuple makes it hard to show the lower level consuming
 Queryperf/DNSperf matches the testing device’s response rhythm to perform fixed
intensity stress test.
 Visit frequency of Internet usually follows Poisson distribution with fluctuation
 Don’t forget human-caused visit frequency fluctuation
 Regular probing, regular web crawler

21. Network I/O selection
Performance
Magnitude
Q40M - 40Gb
Q10M - 10Gb
Q1M - 1Gb
Q100K - 100Mb
Q10K - 10Mb
Q1K - 1Mb
Network I/O
DPDK
PF_RING,Netmap
BPS,netfilter
Socket

22. New DNS data storage
 Open LDAP LMDB
 Lightning Memory-Mapped Database
 PowerDNS + LMDB
 400kqps
 Knot DNS
 700kqps

23. Load Balancer Selection
 Switch/Router + Anycast in the Node
 First choice if condition allows
 Hardware Load Balancer
 Software Load Balancer

24. Intelligent DNS and CDN Traffic Scheduling
 Example analysis
 Factors that impact DNS scheduling
 How to assess DNS scheduling result
 Scheduling alogrithm can be used

25. DNS Scheduling optimizing examples

26. Factors that impack intelligent DNS scheduling
 Intelligent DNS doesn’t interact directly with terminal
 Some terminals change its Local DNS
 Local DNS Cache has impact on scheduling effection and expiration
 Local DNS covers varied sized user groups
 Local DNS optimizing action impacts data equity

27. Accuracy and precision of DNS Scheduling
0%
10%
20%
30%
40%
50%
60%
70%
High accuracy, low precision
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
Low accuracy, high precision
Actual Result Scheduling Target

28. Impacts of accuracy and precision of intelligent DNS
 Precision
 Impact dynamic adjustment of the scheduling system
 Impact device bandwidth usage
 Accuracy
 Impact prediction and planning of the scheduling system

29. DNS Scheduling basis
 DNS statistics data
 Better not use directly
 Intrinsic properties
 Local DNS IP location property
 Local DNS IP netowrk property
 Stateless properties
 Hash characteristic
 Random

30. Common Scheduling Categories
 Default equally divide by traffic
 By propotion
 IP categories
 Static location categories
 Dynamic network link categories
 …

31. Characteristics of Common Scheduling Categories
Category Accuracy Precision
By location Really bad Good
Default equally divide Very good Very good
Random Accurate Not stable
By Local DNS IPs Bad Good

32. Combine intelligent DNS with other scheduling ways
 Intelligent DNS for coarse-grained, other ways for finer grainuality
 Http 302
 Cluster
 Limit connections and traffics on devices
 Mixed uses in node for better productivity according to their own cases
 High quality and small traffic
 High precision intelligent DNS scheduling
 Low demand and huge traffic
 High precision intelligent DNS scheduling and cross-node scheduling
 Limit connections and traffics on auxiliary devices

33. Security
 Software pitfalls
 Heterogeneous software backup for each other
 Open source DNS based
 Function tailoring
 CDN is the most fundamental and stable part of the DNS software
 Useless code branches should be cut out
 DDos traffic attacking
 High performance unusual DNS software can be designed with the bandwidth advantages of
CDN
 Introduce 3rd part DNS security products
 Beware of the impact of security product introducing on traffic scheduling
 Improve the software performance to help the 3rd party security product and lower risks

34. High Availability
 Distribution deployment
 Heterogeneous DNS software
 Platform images

35. Heterogeneous DNS
 Combine two DNS software into one group
 13 global roots of DNS are mixed in ues for BIND and NSD
 Why heterogeneous DNS?
 Using the inconsistency of different DNS’s defects to ensure high availability
 Being one of the critical system, defects in DNS can be catastrophic
 Local DNS cluster retrying can be contagious and paralyze the whole sets of DNS
 Cost of development and maintenance
 DNS is simple and the development cycle is short
 DNS is the relatively stable network protocol, changes in CDN related part even smaller
 CDN related function
 Can be done when stable
 Open source DNS might be short on customized CDN features, but they are good fail-
safe system.

36. DNS Clone Backup
 Facing platform level and outside
malfunctions
 Platform level domain failure
 Domain registration failure
 Upper and top level domain failure
 Security threat
 Spread the risk
 Don’t put all the eggs in one basket
 Evacuation
 Transfer customer to backup platform
when under failure or threats
user1.cnd1.cn
user2.cnd1.cn
user3.cnd1.cn
user4.cnd1.cn
user5.cnd1.cn
user1.cnd2.comu
ser2.cnd2.com
user3.cnd2.com
user4.cnd2.com
user5.cnd2.com
user1
user2
user3
user4
user5
CDNPlatform
1
CDNPlatform
2
Customer

37. Thanks!
Beijing SpeedyCloud Technology Co., Ltd.,
For More Information, please visit: www.speedycloud.cn
You can also send email to: social@speedycloud.cn