Loading presentation...

Present Remotely

Send the link below via email or IM

Copy

Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.

DeleteCancel

Copy of Untitled Prezi

No description
by

Amar Krishna

on 4 June 2014

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of Copy of Untitled Prezi

Some statistics:-

In recent time bandwidth demand of online applications has grown up 110 times since 1995.

6 outta 100 people have internet access in the developing countries.
Access provided using slow satellite or long distance wireless links.
Results in poor quality internet , at times unusable.


Whats the problem?
Previous works (if any)
traffic analysis shows a “strong feedback loop between network performance and user behavior”
In the face of limited bandwidth, the failure rate of uploads is high , discouraging rural residents from contributing to the Internet content and resulting in consumption of largely Western content.
And few more.....
Why Rural Africa(Zambia)
because the increased fiber- optic capacity , coupled with higher-bandwidth, lower- latency technologies such as terrestrial microwave wireless gives hope for improved Internet access in remote areas.
What the authors talk about in this paper.
Studying the performance and usage implications of an Internet Access upgrade.
From 256kpbs->2mbps
No immediate change in usage
Significant change in performance
Subscribers began to use to faster internet eventually.
But sudden increase in usage resulted in bad network performance.
AMAR KRISHNA
Increased B/W fallacy in rural Zambia
Subject location: Macha( Few Facts)
Population -135000
Density- 25 persons/km2
Estimated average income of household is 30 times less than a monthly 1Gb Internet subscription.
An NGO names Linknet provides the Internet services
The lack of electricity coupled with the high prices for user equipment and Internet pro- visioning makes it virtually impossible for Machans to use Internet at home.
2008-2011, internet connection through a satellite connection, approx. cost $1200/month for 256kbps downlink, 64kbps uplink.
04/2011, upgraded to 2Mbps costing $3600/month.
Network Analysis
Duration - 3 months
One month pre-upgrade, one month post-upgrade, to measure the short term impact on n/w usage and performance
Evaluation of one month of traffic approx three months Post-upgrade (Long-term).
Prime focus on network performance analysis, 93% of traversing traffic is TCP.
Describing trends in uplink and downlink performance of TCP flows, and we identify the most popular applications based on TCP port number.
N/w analysis contd.
Assessing the success and failure rate of TCP flows.
Analyzing the “worldliness” of network flows initated in Macha in an effort to determine whether Machans started using more global services once they had better Internet access.
Pre upgrade network- saturated, high RTT, congestion and absorted sessions.
Post upgrade- decreased retransmissions, RTT due to improved network performance.
Post upgrade vs long term- long term users used it for b/w hungry applications such a video streaming.
Methodology
Connecting a monitoring server between Internet gateway and Macha's WLAN
Mirror port to allow capturing of all the traversing traffic.
Metrices used - TCP window size, RTT, TTL and transmission.
Extraction of metrices using tsharks in offline more.
developed a tool that reassembles unidirectional flows from a list of packets based on packet signature( for flow analysis)
Overall N/W performance
Starting with evaluation of traffic load.
Formula used:-
load= Agg. number of bits traversing each hour/number of seconds an hour
Results(Average Traffic load):-
Pre-upgrade-> 367.3 kbps
Post-upgrade->495.3 kbps
long term-> 648.1 kbps
Next fig plots over time the traffic load averaged per hour in blue and the Service Level Agreement (SLA) with the Internet provider in red.
The guaranteed speed was 256kbps, bursts of up to 1Mbps were possible depending on link utilization. This is why the actual traffic load Pre-upgrade consistently exceeds the SLA(service level agreement) of 256kbps.
Contd.
period before the upgrade, the demand frequently exceeded the SLA of 256kbps.
three months after the upgrade the demand often approaches the SLA.
gaps in the plots correspond to time periods in which traffic captures were unavailable due to power or network outages.
Contd.
Total bytes traversed doubled in three months.
Total number of packets dipped post upgrade as same traffic demand was accommodated with fewer transmissions.
RTT trend almost similar.
Immediately after upgrade RTT decreased by 35ms( 3 times as time progressed).
long-tail distribution of RTT in Post-upgrade and Long-term performance

the long-tail distribution of RTT after the upgrade is due to changed browsing habits and tendency to use services that are physically further away
Contd.
Clear bi modality of payload size
Reason being prevalence of either control packets with 0 bytes payload or data packets with payload of about 1500 bytes.
Increase in the average TCP win- dow size (Table 1), which allows more packets to be sent in the network before an acknowledgment is received.
Increased TCP window size is critical to improved TCP performance
Measuring the overhead:- percentage of the total packets that are due to retransmissions and control packets.
The fraction of control packets decreased after the link upgrade from 56.59% to 47.69% and then slightly increased in Long-term to 49.72%.
Contd.
Performance trends over time
This Fig. plots the average on a weekly basis aggregated per hour.
First data point shows an average over all occurrences of the first hour of monday.
clear diurnal pattern in link utilization.
The amount of traffic generated during weekdays differs from that on weekends, with weekday traffic loads typically being heavier.
93% of the traffic traversing the gateway in Macha is TCP.
Bytes in flights?
which is the fraction of sent data that has not yet been acknowledged.
The bytes in flight is influenced by the TCP window size: the better the link performance, the larger the window size.
Contd.
Three periods described here:-Immediately after the upgrade, the bytes in flight drastically increased and continued growing over the Long-term.
Developed a tool to examine the packet signatures of TCP flows through tshark. sourceIP-sourcePORT-destinationIP-destinationPORT.
General trend of TCP flow:-
Total bytes associated with TCP flows increased after the upgrade and continued growing in Long-term.
Due to increasing browsing and use of streaming applications
The flow size doubled Post-upgrade and then continued increasing in Long-term. The increase of flow size can be attributed to different applications utilizing the link immediately after the upgrade and in Long-term.
Uplink & Downlink flows
Aggregate bytes in each direction.
Uplink and Downlink bytes increased after the line upgrade.
Average uplink packet size and flow size increased Post-upgrade and in Long-term.
Downlink packet and flow sizes increased Post-upgrade and then slightly decreased over the Long- term.
Rapid increase in downlink activity Post-upgrade is due to an increase in automated activities such as soft- ware updates.
The number of flows in both up- and downlink directions decreased dramatically Post-upgrade and then increased.
Initial decrease- higher rate of successful flow completion
Subsequent increase in the Long-term is due to a combination of increased user activity
Contd.
Top 10 applications.
HTTP and HTTPS were persistently the most prevalent applications across the three periods.
Pre-upgrade we see many requests associated with Skype and e-mail clients (SMTP, IMAP, POP3).
Post-upgrade security vulnerabilities, such as the one associated with port 8008,disappeared as computers were more successful in downloading and installing critical security updates.
Rest of the ports - Skype and P2P networks.
In terms of fraction of download bytes, HTTP and HTTPS are very high PRE and POST upgrade.
Amount of HTTP bytes decreased by 14% over the long term.
Contd.
Fig. 7(a) presents the fraction of completed and failed flows in uplink and downlink in each period. The completion rate of uplink flows remained unchanged Post-upgrade and then slightly increased in Long-term.
Downlink flow remained unchanged.
Analysis of success and failure trends correlated with byte volume and flow size is done.
Fig. 7(b) plots the aggregate flow size in bytes for each direction.
Post upgrade- remained same ; long term - increased.
The amount of bytes in failed flows approaches the amount of bytes in completed flows.
Downlink flows:- Increase in successful downloads, Long term- nearly reaching the aggregate size of successful downloads.
Average flow size of completed and failed flows in Fig. 7(c).
Uplink:- long term- average size of failed flows is 4x the size of completed flow.
Downlink flows:- Failed D/F higher than completed flows.
Windows vs linux
Using the observed TTL values, the two operating systems were distinguished.
Linux implements CUBIC TCP, which has optimized congestion control mechanisms for high bandwidth networks with high latency.
CUBIC TCP measures congestion independently from long RTTs.
Windows implements TCP Reno in Windows XP and Compound TCP in Windows Vista and subsequent Win- dows versions.
TCP Reno and Compound TCP base win- dow size on the RTT – window size increases with low RTT values and decreases with high RTT values.
Linux maintains higher mean frame size over all three periods.
More aggressive
Throughput in case of linux is 487.6 kbps and in case of windows is 106.2.
Network Usage
6.54% of the Pre-upgrade traffic in up- and downlink direction was a combination of HTTP and HTTPS. This number remained almost unchanged Post- upgrade – 85.85%, and dropped to 67.63% in the Long- term.
Facebook and Google are clearly the most popular sites.
Access to Twitter, the third most popular domain Pre-upgrade, dropped significantly.
Software update sites such as those associated with Win- dows, Adobe, and Ubuntu remain relatively unchanged.
Advertisement-related sites are the seventh most popular request type, representing roughly 2% of all requests.
Contd.
results are divided in terms of flow completion and failure.
Both the number and total bytes associated with Facebook flows increased over the three periods.
This trend is different than the one followed by Google
failure rate of YouTube flows was high Post-upgrade and then decreased.
large flows were most often the ones to fail.
TCP flows from software upgrades:- such flows decreased slightly Post-upgrade and then increased in the Long-term.
Quantity of bytes associated with software updates doubled immediately after the link upgrade.

Contd.
200, 400, and 408. 200 (OK) responses indicate a valid request for which an HTTP server can correctly craft a response.
HTTP 200 responses increases more than 10% after the link upgrade.
400 errors decrease significantly after the link upgrade.
Immediately after the upgrade hosts could have implemented over- due software updates which could rectify browser version issues associated with request format.
anti-virus software was updated to newer versions which could potentially allow for the detection and removal of malware on hosts.
number of 408 er- rors decreases dramatically after the link upgrade.
Contd.
calculating the straight- line distance between Macha and the given coordinates for the other side of each connection using the Haversine formula.
flows generally occur over longer distances in the periods after the network upgrade.
Long-Term flows show even longer distances as compared to Post-upgrade
increase in distances from Macha is the result of a better user-experience after the network upgrade
GeoIP database to find the country code for each external node.
Traffic to and from nodes in Zambia itself increased dramatically after the network upgrade.

each successive period shows an increase in the radius compared to the prior periods.
Network users are connecting to content from a larger geographic variety of the world.
Longer- distance flows in the periods after the network upgrade.
Center- mass values to determine whether the upstream provider changed the distribution.

Contd.
Center-mass points for each period as well as each calculated radius of gyration.
Center-mass values are within a reasonable range of each other given the global scale.
Increase in spread can be credited to an increased geographic diversity of external nodes.
Conclusion
Unique dataset from a rural sub-Saharan village that captures usage before and after an Internet access speed upgrade.
Effect of upgrade on performance and user behaviour.
Difference in performance between Windows and linux machines.
Careful selection of operating system and/or modifications of the network stack to facilitate better networking performance in bandwidth- constrained environments.
In developing rural regions, an Internet upgrade can be just a small increment to the more substantial access speed that is needed to accommodate modern web content and applications.
Such increment gives users the ability to more fully utilize the modern Internet with bandwidth-intensive applications
Full transcript