1 internet performance monitoring update les cottrell & warren matthews – slac presented at...
TRANSCRIPT
1
Internet Performance Monitoring Update
Les Cottrell & Warren Matthews – SLACwww.slac.stanford.edu/grp/scs/net/talk/mon-escc-apr00/
Presented at the ESCC meeting Pleasanton April 26, 2000Partially funded by DOE/MICS Field Work Proposal on Internet End-to-end Performance Monitoring
(IEPM), also supported by IUPAP
3
PingER• Measurements from
– 30 monitors in 15 countries– Over 500 remote hosts– Over 70 countries – Over 2100 monitor-remote site pairs
• Recent monitor additions: ANL, UWisc, NSK, ITEP, RIKEN, KAIST, ILAN, Brazil, Melbourne; working on: Caltech, SDSC
• Over 50% of HENP collaborator sites are explicitly monitored as remote sites by PingER project– Atlas (37%), BaBar (68%), Belle (23%), CDF (73%), CMS (31%),
D0 (60%), LEP (44%), Zeus (35%), PPDG (100%), RHIC(64%)
• Remainder covered by Beacons– Currently 56, extending to 76
4
Beacons & UK seen from ESnet
Sites in UK track one another, so can represent with single site
2 Beacons in UK Indicates common source of congestionIncreased capacity by 155 times in 5 years
Effect of ACLs
Direct peering betweenJANet and ESnet
6
Validations: Ping vs. Surveyor
Scatter plot Ping RTT vs Surveyor RTT gives R2 ~ 0.92www.slac.stanford.edu/comp/net/wan-mon/surveyor-vs-pinger.html
7
RIPE vs Surveyor 1/2
Little short term correlationeven for time differences of< 2 secs
Little structureoutliersdon’t match
9
PingER vs AMP
Little obvious short term agreement (R2<0.1)Same if compare ping vs. ping
Avg Ping distribution agrees with AMPBoth show >=95% of samples are 58-59 msecR2 > 0.95 for min & avg
Time series
10
Rate Limiting 1/3 (Mit Shah)
“Tail-drop” behavior
• Rate-limiting kicks in after the first few packets and hence later packets are more likely to be dropped
Calculate slope and histogram slope frequency for all nodes, look at outliers (8)
Added as PingER metric, Still validating, some sites consistentothers vary from month to month
11
Rate Limiting 2/3Asymmetry of Ping vs Sting losses
-0.60
-0.40
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
clan
2.fit
.uni
mas
.my
ultr
a.he
pi.e
du.g
e
ww
w.d
olph
inic
s.no
ns.u
cr.a
c.cr
cab.
cnea
.gov
.ar
tjev.
tel.f
er.h
r
lhr.
com
sats
.net
.pk
pknt
.utm
.my
tnp.
saha
.ern
et.in
ww
w.ji
nr.d
ubna
.su
ns.it
ep.r
u
gam
ma.
carn
et.h
r
intr
ans.
baku
.az
ns1b
.itb.
ac.id
daim
on.u
nian
des.
edu.
co
groa
.uct
.ac.
za
tifr.
res.
in
ww
w.b
u.ac
.th
ww
w.u
sm.m
y
moo
n.at
omki
.hu
cni.m
d
sun.
ihep
.ac.
cnAs
ym
= (
p-s
)/(p
+s
)Measured 4/22/00 for hosts seen from SLACwith high tail-drop.Hosts selected with > 0.7% loss and no sting pathologies
Hosts mainly in former E. block, S. Asia, Latin America & S. AfricaLarge asymmetry means ping loss >> sting loss, maybe limiting
12
Rate Limiting 3/3• Have identified about 2% of sites possibly limiting • Using Sting (Stefan Savage) & SynAck (SLAC)
tools to identify loss(sting or synack probes) << loss(ping)
• www.vincy.bg.ac.yu blocked 884 rounds of 10 ICMP packets each, out of 903
• islamabad-server2.comsats.net.pk – blocked 554 out of 903
• leonis.nus.edu.sg– blocked all non 56Byte packets
• All low loss with sting or synack
13
Results:How are the U.S.
Nets doing?
In general performance is good (i.e. <= 1%)ESnet holding steadyEdu (vBNS/Abilene) improving, got bad recentlyXIWT (70% .com) 5-10 times worse than ESnet
14
How are DoE funded Edu sites doingEdu seen from ESnet Labs
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Jun-97 Jan-98 Jul-98 Feb-99
Aug-99
Mar-00
Oct-00
Ma
r-0
0 m
ed
ian
% lo
ss
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
Median loss
% sites with > 1% loss
Expon. (Median loss)
Expon. (% sites with >1% loss)
V. poor (> 5% & < 12%):PVAMU, VTechvBNS,
Acceptable (> 1% & < 2.5%): Brandeis, RicevBNS, UCRvBNS, UIUCvBNS (2 bad days in March), TAMUI2
Pairs = 137Fraction NOT good: reduced by 2 in 1.5 yrs
15
Europe seen from U.S.
650ms
200 ms
7% loss10% loss
1% loss
Monitor siteBeacon site (~10% sites)HENP countryNot HENPNot HENP & not monitored
18
Quality of Service: How to improve• More bandwidth
– Keep network load low (< 30%) – Costs (at least in the W) are coming down dramatically,
but non-trivial to keep up
• Reserved/managed bandwidth generally on ATM via PVCs today
• Differentiated services
19
Effect of more & managed bandwidth
German Universities as good as DESY after Oct-99 upgradeDFN closes Perryman POP loses direct ESnet peeringPeering re-established via Dante @ 60 Hudson
RTT
Loss
22
Bulk transfer - Performance TrendsBandwidth TCP < 1460/(RTT * sqrt(loss))
Note: E. Europe not catching up
ESnetFlatteningout
23
Interactive apps - JitterSLAC<=>CERN two-way
instantaneous packet delay variation
0
10
20
30
40
50
60
70
80
90
-100 -8
0
-60
-40
-20 0
20
40
60
80
100
Ping inter packet delay difference in msec.
Fre
qu
en
cy
0
10
20
30
40
50
60
70
80
90
Frequency
Gaussian
Average = -0.03 msec.Std dev = 35 msec.Median = 0 msec.IQR = 29 msecLoss = 0.3%1000 samples
Gaussian-prob=79*exp(-x**2/(2*(IQR/2)**2))
IPDD(i) = RTT(i) - RTT(i-1)
24
SLAC-CERNJitter
IQR(ipdv) between CERN & SLAC from Surveyor measurements (12/15/98 & medians for Dec-98)
0.1
1
10
100
0 5 10 15 20 25
Time since midnight (GMT)
IQR
(IP
DV
) in
ms
ec
.
IQR(ipdv) CERN>SLAC IQR(ipdv) SLAC>CERN
Monthly IQR(ipdv) CERN>SLAC Monthly IQR(ipdv) SLAC>CERN
ITU/TIPHON delayjitter threshold
(75 ms)
25
Voice over IP: Reachability Within N. America, & W. Europe loss, RTT and jitter is acceptable for VoIP
But what about reachability
26
Availability – Outage ProbabailitySurveyor probes randomly 2/secondMeasure time (Outage length) consecutive probes don’t getthrough
http://www-iepm.slac.stanford.edu/monitoring/surveyor/outage.html
27
Error free secondsTypical US phone company objectives are 99.999%
http://www-iepm.slac.stanford.edu/monitoring/surveyor/err-sec.html
What do we see for the Internet using Surveyor measurements
28
• SLAC & LBNL have a DS testbed with a 3.5Mbps ATM PVC carved out of 43Mbps
• Made measurements with Becca Nitzan @ ESnet
Differentiated services & VoIP
PBX
VoIP ESnet
ATM
Bottleneck3.5Mbps
Prod
Edge
WFQ
CAR marking
•Apply WFQ & policing (via CAR)
•With WFQ call sounds fine
–Next use ping to characterize:•Mark ping TOS bits with CAR, & use WFQ in routers and see how it affects loss, RTT, jitter etc.
4Mbps
–Inject 4Mbps UDP load•No WFQ can’t make call
–If make call then terrible quality
–Make phone call–< 50% load call OK
24kbps
29
Plans 1/2• HEPNRC now rejoined at 50% person
• Monitoring – next 2 weeks: select packet sizes, number in stream -
need for• better statistics for high performance links (e.g. PPDG)
• lower impact on low capacity links
– select scheduling, what is logged, mechanism (synack, ping sting)
• Beacons extend from 50 => 70 (requires new mon)
30
Plans 2/2• With XIWT/DARPA
– Anomaly detection and alerting
– NIMI integration
• More graphical reports– Maps, Java servlet graphs of more metrics and more selectability
– Health watch – upper level displays
– Near realtime for SC2000 – possible interest from ESnet NOC• Maps with colored links with playback
• 3D bar charts
• Extended PPDG support– Higher statistics, better coverage
31
Summary• Long term agreement between AMP, PingER,
Surveyor, & RIPE– need persistent structure (e.g. congestion or route
changes) for short term point by point agreement
• Rate limiting still a minor effect, but could become a problem, trying to get good signature, have alternates
• International performance from US to sites outside W. Europe, JP, KR, SG, TW is generally poor to bad
• Managed bandwidth can be big help.• ESnet & Internet 2 doing well, even for VoIP,
except reachability has a way to go
32
More Information• This talk:
– www.slac.stanford.edu/grp/scs/net/talk/mon-escc-apr00/
• IEPM/PingER home site– www-iepm.slac.stanford.edu/
• Comparison of Surveyor & RIPE & PingER– www.slac.stanford.edu/comp/net/wan-mon/surveyor-vs-ripe.html– www.slac.stanford.edu/comp/net/wan-mon/surveyor-vs-pinger.html
• Detecting ICMP Rate Limiting– www.slac.stanford.edu/grp/scs/net/talk/limiting-feb00/