Slide 1: thoughts on measurement and management of the DNS system

thoughts on measurement and 
management of the DNS system




kc claffy
Cooperative Association for Internet Data Analysis (CAIDA)
(UC's San Diego Supercomputer Center)

www.caida.org

Slide 2: before i give this talk

before i give this talk

admit not clear to me what problem we are trying to solve 
DNS as searching and navigation system? ( is that serious? )

assess effect on name assignment, addressing, & searching
of growth in users and sites
of growth in embedded computing devices
of growth in personal and object identifiers

evaluate technologies that can affect Internet searching
addition of generic TLDs,
new name assignment, addressing, and indexing schemes
new directory structures for locating information/sites 
improved user interfaces for accessing info on Internet
navigate trademark, monopoly  hell
evolution: competition; stability; portability

institutions, policies, procedures to implement technologies

Slide 3: instead of giving this talk

instead of giving this talk

we are living on severely borrowed time in 
using the DNS system for Anything At All
much less a directory service
no security
no hierarchy (except 1 thin layer at top with:)
13 points of failure
10 in US
6 in same city (DC), another 4 in California
3 behind military bureaucracies
1 in chapter 11
no authority 
no standard performance evaluation or monitoring
no recourse for underperformance
a technical authority noone trusts
a policy authority noone trusts 
but hey it works (noone's more surprised than we)

Slide 4: where my brain goes when `DNS' and `Internet searching' are in same meme

where my brain goes when `DNS' and `Internet searching' are in same meme



``This is a crude version of a more advanced utility 
that has never been written.''

          -- X-windows xwud(1) man-page




Problems that remain persistently insolvable 
should always be suspected as 
questions asked in the wrong way.     

                      -- Alan Watts

Slide 5: outline of talk i could (will) give

outline of talk i could (will) give


caida macroscopic DNS measurement activities 
-skitter for rssac  (this talk)
-passive measurements of gTLDs/roots from clients (nevil/evi)
-root server traffic analysis (evi)

root name servers: background
rssac project: background
target list 
measurements
high latency destinations
conclusions

Slide 6: in case i get cut off or you lose consciousness

in case i get cut off or you lose conscoiusnesss


upshot relevant to this committee

if you want to assess performance of the DNS system better
or even if you don't
or if you do put another layer of middleware in

--->  don't have 13 points of [root] failure

or if you do

--->  make managing those points integral to the architecture
(management/modeling/modulation/measurement)

Slide 7: root name servers: background

root name servers: background


existing root name servers (listed alphabetically).



highlighted root servers are monitored by CAIDA.

Slide 8: topology mapping project: background

topology mapping project: background


skitter
http://www.caida.org/tools/measurement/skitter
traceroute-like methodology
increments Time-To-Live (TTL)
ICMP echo requests
small (52-bytes) probe packets
slow-paced

probes measure
IP forward path information 
round trip time (RTT) to destination
thousands of destinations

result 
a ton of data (millions of paths per day, for years)
most comphrehensive macroscopic Internet topology data in world
(low bar)

Slide 9: DNS Clients list

DNS Clients list


common list to run on all monitor probes:
combine individual clients lists from all root name servers
stratify routable IPv4 address prefix space

DNS clents list for this study was created in September 2000
49,374 addresses passively collected from root servers
 8,944 addresses from other CAIDA lists

   => cover more than 58,000 prefixes (out of nearly 90,000 in the BGP table)

augmenting list as new data from root servers available

Slide 10: DNS Clients list: characteristics

DNS Clients list: characteristics






"Top Tens" of the DNS Clients list.

Slide 11: DNS Clients list: characteristics

DNS Clients list: characteristics





distribution of destinations in the DNS Clients list by continents

Slide 12: measurements at each monitor

measurements at each monitor


probes DNS Clients list 7-13 times per day

reaches from 33,000 to 36,000 destinations per day
dips on weekends
decreasing by ~2% per month

unique destinations replying per day, march 2001

Slide 13: measurements at each monitor (continued)

measurements at each monitor (continued)


collects between 250,000 and 450,000 RTT values daily 

replies per day, march 2001

Slide 14: high latency destinations (HLD): definitions

high latency destinations (HLD): definitions


consider RTT distributions in each cycle of probes
large diurnal variations in RTT values

RTT is high if above 90th percentile in given cycle (.5-1s)

a destination is high latency on a given day if it had:
         high RTTs in at least half the cycles on 
         all root server monitors

aggregate two 30-day long sets of data:
1 - 30 December 2000
6 March - 4 April 2001

Slide 15: high latency destinations: persistence

high latency destinations: persistence





left peak: random variations in connectivity

right peak: consistently high latency (RTT) destinations

Slide 16: high latency destinations: by origin ASes

high latency destinations: by origin ASes

Slide 17: high latency destinations: by countries

high latency destinations: by countries

Slide 18: high latency destinations: differences between two data sets

high latency destinations: differences between two data sets


number of HLDs in India, Romania and South Africa has decreased by 20%, 36% and 36%, correspondingly.

number of HLDs in Ukraine more than doubled, and in Chile it increased almost 5-fold.

Thailand, Jordan, Georgia, Costa Rica, Brazil and Fiji contributed each more than 1% of the HLD subset in December 2000.

Bangladesh, Turkey, Bulgaria and Nigeria contributed each more than 1% of the HLD subset in March 2000.

Slide 19: high latency destinations: differences between two data sets

high latency destinations: differences between two data sets



December 2000                                         March 2001

Slide 20: High latency destinations: by continents

High latency destinations: by continents


general geographic pattern same in both data sets

number of HLDs:
in Asia decreased slightly
in South Africa increased slightly

largest proportions of HLD (relative to the target list):
Africa
South America
Asia

Slide 21: conclusions

conclusions


topology & performance data scant

need to monitor ALL 13 root servers to minimize bias in identifying high-latency destinations

high latency: last mile bandwidth or topology deficiency?
further examination with other tools to assess cause of the high latency

need to expand to gTLD servers

future root/gTLD server candidate sites should run a monitor for at least 6 months

Slide 22: www.caida.org/publications/presentations/

www.caida.org/publications/presentations/


kc claffy
UCSD/SDSC/CAIDA
kc@caida.org
www.caida.org