Comparative measurements of Internet traffic using cache-triangle

Landau Institute for Theoretical Physics

Chernogolovka, Russia

{sakr,shchur}@landau.ac.ru

Serge A. Krashakov, Lev N. Shchur

Cache-mesh

The problem definition

We can send requests directly or forward to one of the parents.

- Domains, as a rule, are not compact. Therefore static routing via huge access list is practically impossible.- AS based static routing is also hardly to manage- IP-routing is dynamic- Channel loading varies in time

• There is caching proxy-server in Chernogolovka (proxy.chg.ru)• There are several possible parents - FREEnet, RSSI, OSI -

in Moscow; NLANR in USA• We want to minimize user requests latency time.

Fragment of netprobe database193.233.46.0 681/ 682 1.0 1.0 www.chg.ru www-cache.chg.ru www.gnu.org.ru ikia.ru.ircache.net 545.4 6.0 sd.cache.nlanr.net 589.3 12.0 pb.cache.nlanr.net 594.0 9.0193.233.32.0 106/ 106 1.0 2.0 fortran.org.ru itp.ac.ru www.itp.ac.ru ikia.ru.ircache.net 343.7 5.9 sd.cache.nlanr.net 1054.7 12.0193.233.36.0 192/ 192 5.1 4.0 www.issp.ac.ru alpclub.ru ikia.ru.ircache.net 235.7 5.8193.232.212.0 383/ 385 62.5 6.0 www.rssi.ru www.rka.ru ikia.ru.ircache.net ikia.ru.ircache.net 3.8 1.0 pb.cache.nlanr.net 212.1 12.0 sd.cache.nlanr.net 293.1 21.0158.250.9.0 4/ 4 101.5 8.0 www.npi.msu.su optics.npi.msu.su www.grammy.ru ikia.ru.ircache.net 23.1 5.0 sd.cache.nlanr.net 213.0 20.0 bo.cache.nlanr.net 297.0 19.0140.221.9.0 12/ 12 200.2 9.0 www.globus.org www.mcs.anl.gov ftp.mcs.anl.gov uc.cache.nlanr.net 8.0 6.0 pb.cache.nlanr.net 12.2 6.0 bo.cache.nlanr.net 27.3 8.0 sd.cache.nlanr.net 67.0 9.0 ikia.ru.ircache.net 258.2 16.0194.94.42.0 46/ 123 230.1 15.0 www.springer.de science.springer.de link.springer.de uc.cache.nlanr.net 131.5 16.0 sd.cache.nlanr.net 193.6 22.0 ikia.ru.ircache.net 210.2 16.0 pb.cache.nlanr.net 1656.6 13.0198.9.9.0 2/ 2 254.5 10.0 www.nas.nasa.gov sv.cache.nlanr.net 1.2 4.0 uc.cache.nlanr.net 48.0 8.0 pb.cache.nlanr.net 55.5 8.0 sd.cache.nlanr.net 108.0 10.0 ikia.ru.ircache.net 262.5 13.5

Support of ICMP RTT measurements, netprobe database and ICP v2 (with ICP_FLAG_HIT_OBJ) was introducedin Squid-1.1.19 (1997)

Using this algorithm permits:• automatically choosing shortest (RTT-wise) path to the source server• load balancing between alternative communication channels

Standard Squid Modification

tpd <> tod top + tpd <> tod

Proposed Squid modification

Suppose that we must choose between two or more strategies for caching

1. How can the strategies be compared to obtain the conditions under which some are preferable (more efficient, etc)?2. Are the measurements reproducible ?3. If so, what is the accuracy of the measurements?4. Or, how long should the measurements be continied to attain the given accuracy, e.g., 5%?

Main difficulties:• natural characteristics of human activity• evolution of the network• instability of a path from one point (e.g., user) and another (e.g., server)• problem of resolution: Whole Internet? TLD? AS? Host? URL?

Rewind & Replay (january-february 2000)

The ratio of mean speed of the document retrievingfrom the Web for some top-level domains

with and without cache mesh

all at

au by

caco

m de

edu

ee fi fr

gov it jp

net

nl

nu

org p

lse tw ua uk

num

0

1

2

3

4

5

6

7

8

Tra

nsf

er

Speed R

atio

(With

pare

nts

/With

out pare

nts

)

Domain

Cache triangle

Total number of queries to the different TLDduring three weeks of work of the cache-triangle

in symmetric configuration

N = TLD-2.34(5)

10

100

1000

10000

100000

1000000

1E7

Nu

mb

er

of

qu

eri

es

top-level domain

Relative difference in the number of queries served by the Left-slave and

by the Right-slave

ruco

m net

num

org ua de edu nu es uk nl jp err fr

gov cz it

dk ee cc tw lv ch se to cail

by au hu bg be ws sk kz at pl

md kg fi cx br az tv kr us ge am no cn pt gr sg ro ar lu St

nz si cl sh

-10

-5

0

5

10

Asy

mm

etr

y of

the

num

ber

of

que

ries,

%

top-level domain

nl cz ca fi pt

The total traffic served by the Left-slave () and by the Right-slave () for different domains.

N = TLD-2.25(16)

0,1

1

10

100

1000

10000

100000T

ota

l tr

aff

ic,

MB

top-level domain

ruco

m net

num

org ua de edu nu es uk nl jp err fr

gov cz it dk ee cc tw lv ch se to cail

by au hu bg be ws sk kz at pl

md kg fi cx br az tv kr

-100

0

100

Asy

mm

etr

y in

the

tota

l tra

ffic

top-level domain

Relative asymmetry of the total traffic in symmetric cache-triangle configuration

Average document size for the symmetric configuration

ruco

m net

num

org ua de edu nu es uk nl jp err fr

gov cz it

dk ee cc tw lv ch se to cail

by au hu bg be ws sk kz at pl

md kg fi cx br az tv kr us ge am no cn pt gr sg ro ar lu St

nz si cl

0

20

40

60

Left slave Right slave

Ave

rag

e d

ocu

me

nt s

ize

, K

B

top-level domain

Average document transfer speed in kbit/sec, served bysymmetric cache-triangle configuration

ruco

m net

num

org ua de edu nu es uk nl jp err fr

gov cz it

dk ee cc tw lv ch se to cail

by au hu bg be ws sk kz at pl

md kg fi cx br az tv kr us ge am no cn pt gr sg ro ar lu St

nz si cl sh

0

10

20

30

40

50

60

70

aver

age

tran

sfer

spe

ed,

kbps

top-level domain

Left slave Right slave

Average size of the documents served by the Left-slave server using cache-meshand by the Right-slave server with the direct connection to the origins.

ruco

mnu

mne

tor

g ua de nu es uked

ugo

v nl tw err

se fr cz by jp pl beit

dk ca to cc ch sk tv lv fiau kz ee

il at bg md cx ge br ws kr no lu

am hu kg us za mil

az sg int lt

as cn uz ar pt pe gr si

0

10

20

30

40

50

aver

ag

e d

ocum

ent

size

, K

B

top-level domain

Using cach-mesh direct

The average document transfer speed in kbit/sec.

ruco

mnu

m net

org ua de nu es uk

edu

gov nl tw err

se fr cz by jp pl beit

dk ca to cc ch sk tv lv fiau kz ee

il at bg md cx ge br ws kr no lu

am hu kg us za mil

az sg int lt

as cn uz ar pt pe gr si hk

0

10

20

30

aver

age

tran

sfer

spe

ed

top-level domain

Using cache-mesh Direct

The average document transfer speed in kbit/sec.(another data set after tuning configuration)

ruco

mnu

m net

org ua de es uk nu edu nl kr err

gov jp fr se be by tw ca dk it cz pl to il

ee tv cc lv au ch sk at lu kz cx md

am br hu ie ws fi

no hr int

kg bg pe ar tr az sg gr zaS

UM

Non

-RU

0

10

20

30

40

Ave

rage

tran

sfer

spe

ed, k

bps

Top-level domain

Using cache-mesh Direct

Average transfer speed ratio with and withoutusing cache-mesh

ruco

mnu

m net

org ua de es uk nu edu nl kr err

gov jp fr se be by tw ca dkit cz pl to il

ee tv cc lv au ch sk at lu kz cx md

am br hu ie ws fi

no hr int

kg bg pe ar tr az sg gr zaS

UM

Non

-RU

0

5

10

15C

ache

-me

sh/D

irect

Top-level domain

Conclusions

1. Experimental approach for comparative measurements for different cache strategies or cache request routing was proposed.

2. Usefulnes of such measurements for tuning cache server configuration and cache-mesh optimization was shown.