This document summarizes a study on DNS performance and caching effectiveness. It analyzes DNS query logs from MIT networks over three 1-week periods. The study finds that A lookups have a cache hit rate of 80-87%, but caching is less effective due to browsers also caching. It also finds that unanswered lookups and referral loops generate a significant number of retransmissions, accounting for around 60% of all queries. Popular domains have shorter Time To Live (TTL) values that reduce over time.

1. DNS
6 November 2009
CS5229 Semester 1 2009/10
1

2. root
.com
.org
.sg
yahoo
google
pbs
nus
ntu
CS5229 Semester 1 2009/10
2

3. root
server
.sg
.nus
local
DNS
dig +trace www.nus.sg
CS5229 Semester 1 2009/10
3

4. uses UDP
6 November 2009
CS5229 Semester 1 2009/10
4

5. Resource Record"
(name, TTL, class, type, value)"
ns1.nus.edu.sg. 6562 IN A 137.132.123.4
6 November 2009
CS5229 Semester 1 2009/10
5

6. TTL-based Caching
6 November 2009
CS5229 Semester 1 2009/10
6

7. Resource Type"
A NS PTR MX CNAME …"
ns1.nus.edu.sg. 6562 IN A 137.132.123.4
6 November 2009
CS5229 Semester 1 2009/10
7

8. DNS-based "
Load Balancing
6 November 2009
CS5229 Semester 1 2009/10
8

9. DNS-based "
Server Selection
6 November 2009
CS5229 Semester 1 2009/10
9

10. how fast?"
how efficient? "
is caching useful?"
how big should TTL be?
6 November 2009
CS5229 Semester 1 2009/10
10

11. “DNS Performance and
Effectiveness of Caching”"
J Jung et. al."
IEEE TON 2002
6 November 2009
CS5229 Semester 1 2009/10
11

12. Logger
Internet
MIT Subnets
6 November 2009
CS5229 Semester 1 2009/10
12

13. 3 traces"
1 week each
6 November 2009
CS5229 Semester 1 2009/10
13

14. DNS query/respose"
TCP SYN/FIN/RST
6 November 2009
CS5229 Semester 1 2009/10
14

15. basic analysis
6 November 2009
CS5229 Semester 1 2009/10
15

16. DNS Query Types for mit-jan00
6.4
6.8
A
PTR
MX
24.6
ANY
60.4
Others
6 November 2009
CS5229 Semester 1 2009/10
16

17. 496,802"
lookups followed by TCP
connections"
3,619,173"
TCP connections
6 November 2009
CS5229 Semester 1 2009/10
17

18. 80-87%"
DNS cache hit ratio"
for A lookup
6 November 2009
CS5229 Semester 1 2009/10
18

19. not very high since it includes
Web browser’s"
caching too
6 November 2009
CS5229 Semester 1 2009/10
19

20. how fast is DNS lookup?
6 November 2009
CS5229 Semester 1, 2009/10
20

21. lookup latency ≈ "
time between first query"
and last response
6 November 2009
CS5229 Semester 1 2009/10
21

22. mit
kaist
22

23. Number of referals per lookup
mit-jan00
1.16
0.15
24.07
0
1
74.62
2
> 2
23

24. 0
1
2
24

25. NS Cache"
miss:
first query to gTLD or
root servers"
hit:
otherwise
6 November 2009
CS5229 Semester 1 2009/10
25

26. hit
miss
26

27. 2,486,104"
iterative lookups"
6,039,582"
query packets
6 November 2009
CS5229 Semester 1 2009/10
27

28. significant number of
retransmissions
6 November 2009
CS5229 Semester 1 2009/10
28

29. ~24%"
of lookups are unanswered
6 November 2009
CS5229 Semester 1 2009/10
29

30. persistent retransmission"
referral loops
6 November 2009
CS5229 Semester 1 2009/10
30

31. answered
zero referral
31

32. no need to retransmit beyond
2-3 retransmissions!
6 November 2009
CS5229 Semester 1 2009/10
32

33. loops
33

34. each loop generated on average"
10"
query packets
6 November 2009
CS5229 Semester 1 2009/10
34

35. ~60%"
of all queries are generated by
unanswered lookup
6 November 2009
CS5229 Semester 1 2009/10
35

36. popularity of domain name
follows "
Zipf"
distribution
6 November 2009
CS5229 Semester 1 2009/10
36

37. Frequency of the x-th most
popular item is
37

38. popularity of web pages"
population of countries"
occurances of English words"
:"
:
38

39. 39

40. 40

41. 41

42. long tail :"
46%"
of domain names "
are accessed once
42

43. 43

44. 20%"
DNS responses are "
from root/gTLD servers
44

45. without caching NS records:"
5X"
loads on root/gTLD servers
45

46. kaist
mit
46

47. popular sites have shorter TTL"
TLL reduces over time
47

48. 6 November 2009
CS5229 Semester 1 2009/10
48

49. sharing cache does not help
49

50. 50

51. large TTL does not help
51

52. “King: Estimating Latency
between Arbitrary End
Hosts”"
Gummadi et. al."
IMC 2002
6 November 2009
CS5229 Semester 1 2009/10
53

53. how to find alternate path "
with lower latency?
54

54. need to find latency between
pair of Internet hosts
55

55. assume name server "
is close to end host
56

56. authorative
authorative
NS
NS
6 November 2009
CS5229 Semester 1 2009/10
57

57. how to find a name server
close to an end host?"
look at domain name, IP prefix
58

58. how likely does a name server
allow recursive query?"
> 72% of name server tested
59

59. how accurate is the estimated
latency?
60

60. 61

61. 62

62. how close, really, is an end
host to its authorative NS?
63

63. 64

64. 65

65. 66

66. many diverged paths consist of
routers that are physically
adjacent
67

67. 75-80% of the disjoint paths
have latency less than 10ms"
in the Napster trace
68

68. Summary:"
Measurement study of DNS"
Use DNS to do measurement
69