Pakistani Case Study 2010-2011

Deployment

Monitoring Hosts

In 2010, following a series of workshops and site visits, the team at /NUST and SLAC worked with Pakistan’s Education and Research Network (PERN) and Pakistani Universities to put together an end-to-end (E2E) network monitoring infrastructure for PERN connected higher education sites. By the end of 2010 they had have installed the PingER monitoring tools and started gathering data at 18 sites in Pakistan. This includes 4 sites (NUST, COMSATS, PERN and NCP/Quaid-i-Azam) which have been in place for a longer time. In addition they are working on a further 8 monitoring sites. Over 2010 the number of monitoring host – remote host pairs (both in Pakistan) has increased from about 30 to over 220.

The figure below shows the main regions where the hosts are located and how the regions are connected by PERN:

The difficulty of the installations has varied host to host. The technical installation has been simple and not resulted in delays. The delays between deciding to install a monitoring host and gathering measurements from the host have mostly been due to: machine availability; getting administrative approval within university; and delays in making the DNS record entry. The table below shows the history of the installations:

In addition work is in progress to install: BUITMS in Quetta, Bolochistan; Sind Agriculture University Tandojam in Karachi, Sind; NCEMB University in Punjab, Lahore; Virtual University Lahore, Ounjab; Air University in Islamabad, Punjab, Kinnaired College for Women in Lahore, Punjab.

The growth in the number of monitors and host pairs being monitored over 2010, is seen below (spreadsheet):

!pak-infra-2010.jpg|border=1! 

Map of sites

The locations of the Pakistani monitoring (red) and remote(red and blue) hosts are seen in the maps below.

It is interesting to see (right hand map above) the large differences in minimum  RTT between PERN and say NCP (at Quaid i Azam university in N.E. Islamabad) of < 10msec (blue line or more exactly 1.3msec) and that between PERN and NUST in the S.W. corner of between 40-80msec (red line or more exactly 44msec.). Presumably this is due to the routing of PERN connections in Islamabad region. Also interesting to see (middle map above) is that the connection from NUST to the DSL site (blue line) is much less than that to the physically closer NUST host.

Archive, analysis site

IN 2010, a second instance of the SLAC archive-analysis site was set up at NUST. This provides backup for data and access, and improved performance for Pakistani users.

Measurements

Using PingER, the monitoring hosts ping each remote host with 10 pings every 30 minutes. From this data we are able to measure minimum and average Round Trip Times (RTT), jitter, loss, unreachabilty (all 10 pings fail) and derive throughput and Mean Opinion Score (MOS). The data is gathered from the monitoring sites on a daily basis by the archiving  sites at NUST, SLAC and FNAL.

Results

Unreachability

A host is considered unreachable if none of the pings sent to it are responded to.  To illustrate this we chose a reliable host at SLAC  (pinger.slac.stanford.edu) and analyzed the unreachability of Pakistani hosts seen from SLAC.

It is seen that several hosts exhibit high unreachability. The reasons behind the high unreliability are usually site specific and vary from lack of reliable power and a source of backup power, floods, lack of access to the site when there are problems that require physical access, lack of expertise, and lack of interest from a site. 

Mean Opinion Score (MOS)

The telecommunications industry uses the Mean Opinion Score (MOS) as a voice quality metric. The values of the MOS are: 1= bad; 2=poor; 3=fair; 4=good; 5=excellent. A typical range for Voice over IP is 3.5 to 4.2 (see VoIPtroubleshooter.com). In reality, even a perfect connection is impacted by the compression algorithms of the codec, so the highest score most codecs can achieve is in the 4.2 to 4.4 range. Using the RTT, loss and jitter we derive the MOS.

 It is apparent that the MOS is very variable, and according to the middle graph above appears to be decreasing (getting worse) in time (see left hand and middle graphs). Some of this decrease is due to bringing on new hosts that have poorer MOS performance. If we fix on jusr aggregating the performance for hosts pairs that have been monitored for the whole period we get the graph on the right. This set of hosts consists of: PK.NEDUET.EDU.N1, PK.COMSATS.EDU.N2, PK.NCP.EDU.N3, PK.NIIT.EDU.N2, PK.NIIT.EDU.N7, PK.AUP.EDU.N2, PK.PERN.EDU.N1, PK.UET.EDU.N2 and PK.LSE.EDU.N3. In any case the MOS is well above the threshold of 3.5 mentioned above, so VoIP calls within Pakistan between these hosts should be successful.

RTT and Losses for 2010

To try and show the network performance trends within Pakistan in 2010 Amber created a graph showing the inter-regional average RTT performance over the whole year.

Understanding the large variations Jan-Apr 2010

We were concerned about the large variations seen in January 210 through May 2010. Zafar suggested the problem might be to do with the congestion of the links. Jitter (Inter Packet Delay Variation - IPDV) tends to be a good indicator of queuing/congestion, so he suggested looking at this.

The IPDV shows similar behavior to the average RTT:

 
We then looked to see what portion of the larger IPDV early in the year was related to adding new monitoring and remote hosts. If we select the same host-pairs in both say Nov 2010 and April 2010 then the improvement ((ipdv(Apr)-ipdv(Nnov))/ipdv(Apr) in IPDV is about 47%. Thus things have improved with lower IPDVs for the selected host pairs.  

Unfortunately we do not have traceroutes between all the sites to see if they have changed over time. 

Comparing for Nov 2010 the average IPDV for all pairs with that for just the pairs in both April 2010 and November 2010 one gets 5.13ms verses 4.58. Thus some of the reduction appears to be due to the new hosts added.

, prior to this there are less than 100. It is suspected that this is partially the cause of the more stable values of IPDV following this date.

Returning to the average RTT graph, we believe it is less confusing to mainly focus on the data from May 2010 onwards, in order to remove the effects of small statistics. In this case, the average RTTs from Jun through November 2010 have been pretty flat.

Differences in RTTs for regions

Looking at the data for November 2010 it is seen that the average RTts to Karachi and Lahore are two to three times less than those to Quetta, Islamabad and Peshawar. This needs to be investigated.