...

OrgName:    GoDaddy.com, Inc.
OrgID:      GODAD
Address:    14455 N Hayden Road
Address:    Suite 226
City:          Scottsdale
StateProv:  AZ
PostalCode: 85260
Country:    US

Locating PingER Sites

  Using TULIP2 we located all PingER sites in North America. The results can be found here We conducted Tests both using alpha=100km/ms and 40km/ms Currently we are comparing the two results to see, where and how much improvement we are getting.

Alpha Comparison ( 100 km/ms vs 40 km/ms)

 A series of three tests were performed to compare the value of 100 vs 40 for the value of alpha.

• The first test was performed on 3rd April,2008. The results of this test can be found here. A total of 121 North American PingER sites were used in the test.The average error distance for alpha = 100 (km/ms) was 307 km and alpha = 40 (km/ms) was 300 Km. Frequency analysis of sites revealed that some the most frequent sites used were:       

  Landmarks	Frequency
  Chicago_UnitedStates_PL	34
  NewYork_UnitedStates_PL	31
  FNAL,Chicago	26
  Philadelphia_UnitedStates_PL	26
  Bloomington_UnitedStates_PL	22
  Fairfax_UnitedStates_PL	11
  Washington_UnitedStates_PL	11

    
• The second test was conducted on 24th April,2008. The results of this test can be found here. The average error distance for alpha  100 (km/ms) was 238 km and the median value was 243km. For alpha = 40 km/ms the average error distance was 273 km and the median error distance was 183 km.
• Third test was conducted on 1st may,2008. The results of this test can be found here.
  

Adding more precision to TULIP results

   Previously the TULIP algorithm located a host only giving latitude and longitude as integers without any floating points, as a result a lot of precision was lost. Therefore the TULIP algorithm was improved to give floating point precision. In order to test the new accuracy some tests were conducted.

• The first test was conducted on 1st July,2008. Here are the results of the test. Here we compared only those results where the same landmarks were used.
• The second test was conducted on 2nd July,2008. Here are the results of the test. Some quick observations from the second test are:

1. The results are very bad as compared to the previous ones because a lot of landmarks are failing. We are currently thinking on ways to improve on that.
2. The maximum improvement using floating point was 124 Km.
3. Stanford_UnitedStates_PL is giving funny results, will have look deeply into it to find out the cause.

Comparing results with old results after careful selection of landmarks

After careful selection of landmarks (removing non-responding landmarks and writing scripts to dynamically generate landmarks) on 20th Aug,1008 we obtained much better results. These results are compared to the old results in order to see the improvements. The results can be download here Here are some quick observations:

  1. Menlo Park Landmark (Stanford) is giving wrong result (rtt 2ms) which is ruining the results. We need to disable this landmark since we already have landmark at SLAC.

  2. There are some major improvements in results due to careful selection of landmarks.

  3. the median error is 170km, 25% 100km, 75% 362km & IQR 262km. This is including the numbers in red.

  4. Need to come up with a method to avoid bad results for replicated servers.

Abnormal Cases:

 Satellite Links

 Hosts that are on satellite links have very large delays. They cannot be located using the RTTs from a set of landmarks. Usually they have RTTs > 500ms. Most of the Central Asian countries (including Afghanistan) and some African countries use satellite links. Data from hosts in Kazakhistan, Burundi, Madagascar, Uganda, Afghanistan and Nigeria can be found here. The three best Landmarks chosen gives an RTT greater than 500ms. If the RTT from the three best landmarks is greater than 500ms then our algorithm returns the host as satellite link and does not perform trilateration. This figure may be changed latter.

Replicated Servers

 Trilateration VS Multilateration

  Here is the data.