2.2 F AST algorithm and its stability
3.1.1 Burstiness in the paket loss proess
3.1.1.1 Measurement
Generalburstinessinthepaketlossproessesisalsowell-doumented[31,41,42℄.
However, in our assumption, we further laim that the loss proess is bursty in sub-
RTT timesale. We support this assumption with evidene from our measurements
in NS-2simulation,Dummynet emulation,and PlanetLab.
Based on these two assumptions, we model the loss synhronization rate as the
detetion probabilityusing one on-oproess (TCP data pakets) tosampleanother
on-oproess (paketloss proess).
Our model predits that the ombination of bursty TCP ows and a drop-tail
router(bursty lossproess)yieldsverylowanduniform synhronizationratesamong
TCP ows with dierent ongestion window sizes and leads topoor fairness onver-
gene. Our model also suggests that the use of paing at the TCP soures and/or
the use of random dropping algorithms in the link (e.g. RED [31℄) an inrease
synhronization rate.
lost pakets, alled the lossinterval, and analyzed the loss proesses by plotting the
umulative distribution funtion (CDF) and the probability density funtion (PDF)
of thelossintervals. Weompared thePDF ofthe paketlossproessestothe orre-
sponding Poisson proesses with the same average event arrival rates. We observed
that the paketloss proesses are muh burstierthan the Poisson proesses.
The measurements from NS-2, Dummynet, and the Internet all suggest that the
sub-RTT paket lossproess isvery bursty.
Results in NS-2 Simulation Figure 3.1 shows the CDF of the loss interval in
NS-2 simulations. The RTTs of the ows insimulation are random between 2ms to
200ms. From the gure, we observed that 80% of the paket losses luster within
short time periods smallerthan 1% of the RTT.
We alsoplotted the PDF of the loss intervaland ompared it with the PDF of a
Poisson proess with the same arrival rate, asshown in Figure3.1 (B) .
Figure 3.1 (C) zooms in to a small time sale of 0 to 2 RTT and uses log-sale
in the Y-axle so that the Poisson proess has a straight line in its PDF. Compared
to the Poisson proess, the loss proess is muh burstier more than 10 times the
paket losses ourred inthe very smalltime interval.
Results in Emulation Network Figure 3.2 is the CDF of the loss interval in
Dummynet emulations. The RTTs of the ows are xed to 4 lasses: 2ms, 10ms,
50ms,and 200ms. The lossintervalCDF shows asimilarpatterntotheNS-2results,
exept that the CDF starts from 0.1% of RTT due to the limited time resolution of
our measurements inthe Dummynet router.
Figure3.2 (B) and (C)showthe PDF ofthe lossinterval. Again, the lossproess
is muhburstier than the orresponding Poisson proess.
Results in the Internet Figure 3.3is the CDF with the Internet measurement.
TheInternetmeasurementshowslessburstinessinlossproessesthanweobserved
10 −6 10 −4 10 −2 10 0 10 2 10 4 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Loss Interval (RTT)
CDF
Measured
(A)CDF
0 2 4 6 8 10
10 −6 10 −5 10 −4 10 −3 10 −2 10 −1 10 0
Loss Interval (RTT)
Measured Poisson
(B) PDF (Binsize 0.1RTT)
0 0.5 1 1.5 2
10 −6 10 −5 10 −4 10 −3 10 −2 10 −1 10 0
Loss Interval (RTT)
Measured Poisson
(C) PDFEnlarged (Binsize 0.02 RTT)
Figure3.1: Loss intervalsin NS-2measurements.
Note that all the CDF gures in this hapter have X-axles in log-sale, and all the
PDF guresin this thesis haveY-axles inlog-sale.
10 −6 10 −4 10 −2 10 0 10 2 10 4 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Loss Interval (RTT)
CDF
Measured
(A)CDF
0 2 4 6 8 10
10 −6 10 −5 10 −4 10 −3 10 −2 10 −1 10 0
Loss Interval (RTT)
Measured Poisson
(B) PDF (Binsize 0.1RTT)
0 0.5 1 1.5 2
10 −6 10 −5 10 −4 10 −3 10 −2 10 −1 10 0
Loss Interval (RTT)
Measured Poisson
(C) PDFEnlarged (Binsize 0.02 RTT)
Figure3.2: Loss intervals inDummynet measurements.
10 −6 10 −4 10 −2 10 0 10 2 10 4 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Loss Interval (RTT)
CDF
Measured
(A)CDF
0 2 4 6 8 10
10 −6 10 −5 10 −4 10 −3 10 −2 10 −1 10 0
Loss Interval (RTT)
Measured Poisson
(B) PDF (Binsize 0.1RTT)
0 0.5 1 1.5 2
10 −6 10 −5 10 −4 10 −3 10 −2 10 −1 10 0
Loss Interval (RTT)
Measured Poisson
(C) Loss intervalsin PlanetLabmeasurements (Bin size 0.02 RTT)
Figure 3.3: Loss intervals inPlanetLab measurements.
ofappliationtypes, trapatterns,and queuingdelay. Insuhanextremelyhetero-
geneous environment,we observed that 60% of the paket losses lusterwithin short
timeperiodsof1RTT,and40%ofthepaketlosseslusterwithintimeperiodsof1%
of RTT. This evideneis stillvery strongfor sub-RTT burstiness inlossproesses.
Weplottedthe PDFin Figure3.3(B)(C) andompared the Internet lossproess
againsta Poisson proess with the same arrival rate. We observed similar burstiness
as in NS-2 and Dummynet. In the smallest interval region (left side), the measured
lossproess isfar burstierthan the Poisson proess.
3.1.1.2 Possible Soures of sub-RTT Burstiness
As shown by the results of the NS-2 simulations, Dummynet emulations and the
Internet measurements,paketlossishighlybursty insub-RTTtimesale. There are
several possible souresthat lead to suhburstiness.
DropTailroutersare onsideredthe major soureofpaket lossburstiness [31℄. A
DropTailrouter serves asa FIFOqueue, aepting inomingpakets untilthe buer
isfull. Working withDropTailrouters,loss-based ongestionontrolalgorithmskeep
inreasingthe datarate whenthe router'sbuerisnot full. Whenthe router'sbuer
is full and pakets are dropped, the aggregate data rate is higher than the router's
apaity and paket loss persists until the loss-based ongestion ontrol algorithms
detet the loss of pakets and redue the data rate, usually one half of an RTT
later. In between the rst paket loss and the redution of data rate, there is a
peak of paket losses in the DropTail router. Some researhers propose introduing
randomness in the router. For example, Floyd and Jaobson proposed to randomly
dropthe pakets earlierbeforethe buerisoverowed [31℄. However, theseproposals
suer from diultparameter settings problems.
Slow start of TCP ows is another soure of paket lossburstiness. A TCP ow
starts with averysmallrate inburst (sending two pakets bak-to-bak every round
trip), and doubles its data rate if no loss is observed. This proess an quikly
inrease the queuesize inthe bottlenek buer injustafewround tripsand produe
burst period of Flow i
burst period of loss signal randomly drop from M
incoming packets Legend:
a dropped packet a packet
from flow i
i i
i i i i
i i i i i
S incoming packets during the RTT of loss event
i
a packet (from any flow)
spanning over K incoming packets
Figure3.4: Congestiondetetionwithinone RTT:aowuses itsdatapaketproess
to samplethe lossproess. The loss synhronization rate is the probabilitythat one
of the
w i
pakets fromowi
(distributed overKpakets) happens tobeone ofthe L dropped pakets (distributed over M pakets).algorithms, suh as QuikStart [46℄ and RCP [47℄ have been proposed to avoidsuh
aggressivedetetion. Thesealgorithmsrequirehangesindatapaketformats, whih
are expensive for the existing infrastruture.
Hene, thesoures ofsub-RTTburstiness inpaketlossproesseswillexistinthe
foreseeable future.