Upload
lerato
View
70
Download
0
Tags:
Embed Size (px)
DESCRIPTION
PingER : Actively measuring the worldwide Internet’s end-to-end performance. Les Cottrell SLAC MYREN meeting, Kuala Lumpur December 12, 2012. Agenda. Using PingER measurements going back to 1998 and covering 168 countries, this talk will illustrate, Internet performance worldwide. - PowerPoint PPT Presentation
Citation preview
PingER: Actively measuring the worldwide Internet’s end-to-end performance
Les CottrellSLAC
MYREN meeting, Kuala Lumpur
December 12, 2012
AgendaUsing PingER measurements going back to 1998 and covering 168 countries, this talk will illustrate, Internet performance worldwide.
•Brief history
•How can the Internet help development?
•How does PingER measure Internet performance?
•What do we measure, what does it tell us?
•What do we find?
•Case studies illustrating PingER
History• Story of Ping
• Early PingER
• Extension to Developing Regions
• Extension to Pakistan
• Extension to Malaysia
The start• Ping tool invented by Mike Muus
– “a little thousand-line hack” during a single evening to troubleshoot “odd behavior” on the computer network at the U.S. Army’s Ballistic Research Laboratory in Maryland.
– sent a small data packet known as an echo request to an IP address, typically a remote server or network node. If the target address was reachable, it echoed back the same data, and the program recorded the time it took for the round-trip journey.
– Reminded Muuss of the percussive sound pulse sonar systems use to detect objects underwater, he named it after that sound—ping.
5 Joint Techs: I2 & ESnet,Stanford
Measurement Mechanism: PingER
Internet
10 ping request packets each 30 mins
RemoteHost(typicallyweb server)
>ping remhost
Ping response packets
Measure Round Trip Time & Loss
Monitor Host
6
Early PingER• As the head of networking at SLAC, I set up the
system using ping simply to test connections between the laboratory and several dozen research institutions in about a dozen countries that were collaborating on a physics experiment known as BaBar to study properties of subatomic particles.
• Over the next half-decade, as word of PingER’s value spread, I extended monitoring to hundreds more physics laboratories and science centers across the globe. But the project didn’t take a humanitarian turn until 2001.
•UNIMAS
Workshop, Sarawak, Dec 2012
7
Extension to the Developing regions• In 2001 I visited ICTP in Italy.• Driven by ICTP’s goals of bringing first-class science
and technology to developing countries they wanted to know how well the networks were working.
• The simple PingER project was the perfect tool for the job. Ubiquitous ping so nothing to install at remote targets.
• They offered to help expand the project to those parts of the world that needed it most.
• Within the next year, we began establishing monitoring and target hosts in countries as diverse as Ecuador, Rwanda, Jordan, and Bhutan.
UNIMASWorkshop, Sarawak, Dec 2012
8
Extension to Pakistan• In 2004 set up joint agreement with NUST in Pakistan• Soon got my first real glimpse of just how much of a
difference PingER can make. – Set up a PingER monitoring site in the country to
assess performance on the then year-old Pakistan Educational Research Network (PERN).
– The network’s providers touted its bandwidth of 155 Mbps, impressive at the time. But PingER revealed that the “last mile” links to universities were dreadful. These bottleneck connections funneled data at no more than 1 Mbps, causing long delays and high packet loss.
UNIMASWorkshop, Sarawak, Dec 2012
9
Extension to Pakistan
• He clearly took PingER’s lessons to heart. When construction of PERN2 began in 2009, its plans included extending high-speed, 1-Gbps data links all the way to university data centers
UNIMASWorkshop, Sarawak, Dec 2012
• During a visit to the university, I presented our findings to the chairman of Pakistan’s higher education commission, Atta-ur-Rehman, who was preparing to fund the next major upgrade to PERN.
10
Extension to Malaysia
• Have set up an official signed MoU between SLAC & U of Malaysia in Sarawak (UNIMAS)
• Idea was to replicate the NUST project
• Fortnightly meetings by Skype
• Just getting started, no students yet
• Met with Vice Chancellor (VC) at one meeting
• This workshop is a follow up.
UNIMASWorkshop, Sarawak, Dec 2012
11
Why do measurements of the Internet matter
UNIMASWorkshop, Sarawak, Dec 2012
Why does it matter
• African scientists isolated
• Lack critical mass
• Need network to collaborate but it is terrible
• Brain drain
• Brain gain, tap diaspora
• Blend in distance learning
• Provide leadership, train trainers
12
Internet Users 2002
Cartograms from:www.geog.qmw.ac.uk/gbhgis/conference/cartogram.html
Tertiary Education fromhttp://www.worldmapper.org/
13 eGY Africa 2012Workshop, Nairobi Oct 2012
How does the Internet help• Investment in information technology plays the role of
a "facilitator" that allows other innovations to take place. http://findarticles.com/p/articles/mi_m1093/is_3_45/ai_86517828/
• World Bank / IFC report: for every 10% increase in high-speed Internet connections there is an increase in economic growth of 1.3 percentage points. April 2010. http://www.infodev.org/en/Article.522.html
• Example: Uganda 15% increase in price of maize based on improved farmer bargaining power. www.itu.int/ITU-D/.../S1-01-NG-ICT_Indicators-Tim_Kelly.pptx
14
How does PingER work
• Mechanism
• Coverage
UNIMASWorkshop, Sarawak, Dec 2012
15 Joint Techs: I2 & ESnet,Stanford
Measurement Mechanism: PingER
Internet
10 ping request packets each 30 mins
RemoteHost(typicallyweb server)
>ping remhost
Ping response packets
Measure Round Trip Time & Loss
On
ce a Day
Uses ubiquitous ping
Monitor Host
Repositories
NUST
16
Deployment• Monitors > 90 in 23 countries, 4 in Africa
Summer Joint TechsStanford, July 2012
• Beacons monitored by most monitors (~100)• Remote sites monitored by some monitors (~750)
17
Metrics Available from PingER• UnReachability
• Minimum RTT
• Average RTT
• Jitter
• Loss
• Derived throughput
• MOS
• Directness of Connection
• OthersUNIMAS
Workshop, Sarawak, Dec 2012
18 eGY Africa 2012Workshop, Nairobi Oct 2012
Unreachability
Summer Joint TechsStanford, July 2012
Unreachability: e.g. N. African uprisings Jan ‘11
NARSS (Cairo)
Helwan (Cairo)
EUN (Cairo)
23:59 Jan 28
23:59 Jan 27
12:00 Jan 27
• Impact varied: start time, recovery time, after effects• Egypt University Network (EUN) down least time
– NARSS via Alternet->Italy->Egypt, Helwan &EUN via PCCW Global
• Libya first went dark 06:00 Feb 19 for 3 days, then again on Mar 4th more permanently
• Algeria, Morocco, Tripoli not noticeable
=No pings respond
19
Average Round Trip Time (RTT)• Mainly a distance related, but also congestion (i.e. at
the edge)
• For real-time multimedia (H.323) traffic RTT: 0-300ms =Good, 300-600ms=Accceptable, and > 600ms= poor.
UNIMASWorkshop, Sarawak, Dec 2012
20
Minimum RTT history by region• Minimize effects of congestion and queuing
UNIMASWorkshop, Sarawak, Dec 2012
21 UNIMASWorkshop, Sarawak, Dec 2012
•GEOS (Geostationary Earth Orbit Satellite)–Good coverage, but expensive in $/Mbps–& long delays min RTT >450ms easy to spot
N.b. RTTs > 250 ms bad for VoIP
Impact of GEOS vs TerrestrialGEOSGEOS
22 UNIMASWorkshop, Sarawak, Dec 2012
Conversion history by country seen by min-RTT
23
Jitter• Mainly at edges, critical for real time: VoIP, gaming• Exponential improvement (factor 10 in 6 yrs)• The optimum amount of one way latency is 11 ms for keeping time in
music. – Above that delay and they tend to slow down. – >50-70 ms performances tended to completely fall apart.
UNIMASWorkshop, Sarawak, Dec 2012
• For real time haptic control and feedback for medical operations <=80ms is needed.
N. America, Europe, E Asia & Oceania < 1msAfrica, S. Asia & S.E. Asia worst off
24 Summer Joint TechsStanford, July 2012
Losses• Low (<1%) losses are good.• Real time impact due to recovery timeouts, e.g. echoing typing• Losses are mainly at the edge, so distance independent• Losses improving roughly exponentially, ~factor 100 in 12 years
24
Loss has Similar
behavior to thruput
• Best <0.1%: N. America, E. Asia, Europe, Australasia
• Worst> 1%:• Africa & C. Asia
Loss
25 UNIMASWorkshop, Sarawak, Dec 2012
Summer Joint TechsStanford, July 2012
Derived ThroughputDerived throughput ~ 8*1460/(RTT*sqrt(loss))
Mathis et. al
Europe, E. Asia & Australasia merging
Behind Europe:
4 yrs: Russia, 7 yrs:L America,
M East, SE Asia
11 yrs: India, C. Asia
13 yrs: Africa
26
• ITU metric, based on quality of a conversation– Originally people listen and give quality 1-5– Can derive from RTT, jitter and loss
UNIMASWorkshop, Sarawak, Dec 2012
MOSMean Opinion Score MOS)
• >=4 is good,
• 3-4 is fair,
• 2-3 is poor.
Important for VoIP
Usa
ble
From the PingER project http://www-iepm.slac.stanford.edu/pinger
27 UNIMASWorkshop, Sarawak, Dec 2012
Directness of connection (Alpha)• Alpha to allow for delays in network equipment &
indirectness of actual route. D = 1 way distance– Alpha = D(km) / (min_RTT[msec] * 100 [km/msec])
• If know lat/long of monitor and remote host then know D, so with min-RTT can estimate Alpha– Max(Alpha) =1 = direct (great circle) route and no
network delays– Alpha > 1 probably identifies bad lat/long
coordinates for hosts.– Low values typically mean very indirect route, or
satellite or slow connection (e.g. wireless)– Alpha typically ~ 0.45
28 UNIMASWorkshop, Sarawak, Dec 2012
Alpha worldwide
• Interest in Polar route with Global warming
Summer Joint TechsStanford, July 2012
Alpha=0.71A
lpha
=0.7
3
Alpha
=0.7
3SLAC
JP
AU
NZ
NSK.RU
TW0.180.16IN
DE
EG0.34
JP0.32
o.41
AU0.53
29
• Big improvements for C Asia, S Asia & Australasia
UNIMASWorkshop, Sarawak, Dec 2012
Directivity (Alpha) from SLAC to world
Summer Joint TechsStanford, July 2012
More stable year to year as add more hosts
30 UNIMASWorkshop, Sarawak, Dec 2012
Other metrics• Duplicate packets (try ping www.cern.ch, load
balancing?)
• Out of order packets (parallel paths)
• Conditional Loss Probability (non-random loss)
– one packet is lost the following packet is also lost– route change, loss of sync, spanning tree reconfig
• Maximum packet loss (useful for buffer bloat?)
Summer Joint TechsStanford, July 2012
31 UNIMASWorkshop, Sarawak, Dec 2012
More Information
• PingER web home page– http://www-iepm.slac.stanford.edu/pinger/
• Tutorial on network monitoring– http://www.slac.stanford.edu/comp/net/wan-mon/tutorial.html
• PingER data Explorer– www-iepm.slac.stanford.edu/pinger/explorer.html
• Telegeography submarine cable map– http://www.submarinecablemap.com/
32
S.E. Asia
• Just started mining early data
• Where does Malaysia sit
• How much variation in SE Asia
• Variation in Malaysia
• Troubles at UNIMAS
• Top and bottom 3 sites monitored in Malaysia
UNIMASWorkshop, Sarawak, Dec 2012
33
Malaysia vs Other Regions
UNIMASWorkshop, Sarawak, Dec 2012
34
Variation between SE Asian countries
• Factor of 10 between Singapore and Laos
• Singapore 4x better than next countries
• Exponentially improving with time
• On its ownSingaporeapproachesE Asia.
UNIMASWorkshop, Sarawak, Dec 2012
35
Avg-RTT, jitter & loss by Malay State• Need low values of all 3 metrics
• RTTs similar, big diffs in jitter & loss
• Allianze UniColl looks bad
• UTP next worst loss
UNIMASWorkshop, Sarawak, Dec 2012
Seen from SLAC, Nov 2012
36
Non lossy Malaysian hosts seen from SLAC Nov 2012
UNIMASWorkshop, Sarawak, Dec 2012
UNIMAS
MIMOS
MIMOS
UNISZA
Note monitoring host (SLAC) down
37
OCESB
Lo
ssy
ho
st s
een
fro
m S
LA
C N
ov
2012
UNIMASWorkshop, Sarawak, Dec 2012
UPSI
UTEM
MIU
AIU
Sabah
38
Diving deeper: packet loss Nov 27-28• Allianze University College unreachable• UTEM, MIU, UPSI, OCESB, AIU, SABAH experienced loss Nov
26• No PingER loss from the rest on Nov 26: Johor, Kelantan, KL,
Negeri Sembian, Sarawak, Terenggan
UNIMASWorkshop, Sarawak, Dec 2012
1930-midnight MST, backup?
UTEMLosses isolated, not correlated with large RTT
39
Another host, large RTTs correlate with time of day (note MST)
UNIMASWorkshop, Sarawak, Dec 2012
40
UNIMAS Jitter
UNIMASWorkshop, Sarawak, Dec 2012
UNIMAS to
MalaysiaJitter
AllianzeUniversity
College
Universiti TechnologiPetronas
41
Why not show UNIMAS to Malaysia more
• Big changes in RTT affect throughput especially for Kuching
• UNIMAS was seeing congestion – This would be seen everywhere– Turn on shaping– Removes loss &
day-night variations
UNIMASWorkshop, Sarawak, Dec 2012
2
1RT
T m
s
Background loss colors
SLAC to UNIMAS Oct-Nov 2012
42
Increase of capacity to UNIMAS from 200Mbps to 500Mps
UNIMASWorkshop, Sarawak, Dec 2012
43
Improvement
UNIMASWorkshop, Sarawak, Dec 2012
44
Unreachable Malaysian hosts
• Unreachable from SLAC Nov 1-26, 2012:1. 92% Allianze University College
2. 20% www.ocesb.com.my (Speedtest)
3. 17% Universiti Teknologi Petronas,Bandar Seri Iskandar
4. 4% University Malaysia Kelantan
• 100% Reachable– MIMOS, UNIMAS, Sultan Zainal Abidin University,
USIM, University Teknologi Malaysia, Sultan Idris University of Education
UNIMASWorkshop, Sarawak, Dec 2012
45 UNIMASWorkshop, Sarawak, Dec 2012
Demo
• Interactive demonstrations of the data mining capabilities of public data sources provided by organizations such as the UN and ITU coupled with monitoring data from PingER
• http://www-iepm.slac.stanford.edu/pinger/explorer.html
Summer Joint TechsStanford, July 2012
46
Comparison to UN ITU ICT Development Index
UNIMASWorkshop, Sarawak, Dec 2012
100
1000
10000
ICT Development Index from the UN International Telecommunications Union
2 4 6 810
Pin
gER
Der
ived
thr
ough
put
(kbi
ts/s
ec)
Bubble size = Population
47
More Information
• PingER web home page– http://www-iepm.slac.stanford.edu/pinger/
• Tutorial on network monitoring– http://www.slac.stanford.edu/comp/net/wan-mon/tutorial.html
• PingER data Explorer– www-iepm.slac.stanford.edu/pinger/explorer.html
• Invitation to set up a Monitoting host– www-iepm.slac.stanford.edu/pinger/letters/invite-
monitor.doc
UNIMASWorkshop, Sarawak, Dec 2012
48
Congestion at UNIMAS• SLAC to UNIMAS Nov 1st through Dec 6th
• Note the diurnal changes until shaping
UNIMASWorkshop, Sarawak, Dec 2012
Loss
RTT