FLAME Flexible Lightweight Active Measurement Environment

Transcription

1 FLAME Flexible Lightweight Active Measurement Environment Marcos L. Kirszenblatt Thiago B. Cardozo Artur Ziviani Antônio Tadeu A. Gomes National Laboratory for Scientific Computing (LNCC) M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

2 1 Motivation Goal Related Work 2 3 Why Lua? LAMP API 4 Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation 5 Conclusions Future work Questions? M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

3 Motivation Goal Related Work Typical research fronts in active measurements development or improvement of measurement tools deployment of large-scale platforms Problems in developing/improving active measurement tools large development and testing time low code reuse absence of standardized storage of results M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

4 Goal Motivation Goal Related Work A platform for the rapid prototyping of active measurement tools that: offers basic primitives for active measurement Active measurement: involves the sending of probe packets is extensible stores the results in a centralized repository allows carrying out cooperative and uncooperative experiments Cooperative experiment: the tool must be executed at both ends Uncooperative experiment: the tool is executed only at the origin M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

5 Tools for active network measurement Motivation Goal Related Work each tool targets a limited set of metrics low code reuse results are presented in an ad hoc way Tools COP OWD OWV RTT PLR PRO ROT PHC ECP ABW Iperf owping pathchar pathload pathrate ping QoSMet sprobe sting traceroute traceroute-paris Legend: COP: tools that depend on cooperative destination nodes. OWD: one-way delay; OWV: OWD variation; RTT: round-trip time; PLR: packet loss rate; PRO: packet reordering; ROT: route tracing; PHC: per-hop capactity; ECP: end-to-end link capacity; ABW: available bandwidth. M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

6 Platforms for active network measurement Motivation Goal Related Work Manager-agent architecture NIMI ETOMIC DIMES NetQuest Scriptroute is the closest to ours Ruby scripts no centralized repository no cooperative experiments M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

7 XMPP (Extensible Messaging and Presence Protocol) Measurement manager User console Measurement agents Components XMPP server measurement manager user console measurement agents M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

8 XMPP (Extensible Messaging and Presence Protocol) Measurement manager User console Measurement agents XMPP server acts as a message hub among other FLAME components presence control connectivity control offline message store for temporally disconnected components NAT/firewall friendly M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

9 XMPP (Extensible Messaging and Presence Protocol) Measurement manager User console Measurement agents Measurement manager controls the starting and the finishing of experiments only component that accesses the data repository M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

10 XMPP (Extensible Messaging and Presence Protocol) Measurement manager User console Measurement agents User console registers an experiment at the manager allows access to specific agents sends commands or Lua scripts to the selected agent(s) M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

11 XMPP (Extensible Messaging and Presence Protocol) Measurement manager User console Measurement agents Measurement agents host a Lua interpreter run the scripts in a Sandbox environment provide access to a measurement API (LAMP) send the results to the manager may operate in exclusive mode M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

12 Why Lua? Why Lua? LAMP API Lua characteristics interpreted programming language simple syntax extensible semantics portability low memory footprint Don t know Lua? Try it: Lua is a powerful, fast, lightweight, embeddable scripting language. Lua combines simple procedural syntax with powerful data description constructs based on associative arrays and extensible semantics. Lua is dynamically typed, runs by interpreting bytecode for a register-based virtual machine, and has automatic memory management with incremental garbage collection, making it ideal for configuration, scripting, and rapid prototyping. M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

13 Why Lua? LAMP API LAMP API: Lua Active Measurements Primitives The active measurement primitives are implemented in C, being available through a minimalist API LAMP API lamp.sendudpow ( ) lamp.sendudptw ( ) lamp.sendtcpow ( ) lamp.sendtcptw ( ) lamp.sendicmptw ( ) lamp.sendudppt ( ) Notation OW = One Way (Cooperative) TW = Two Way (Uncooperative) PT = Packet Train (Cooperative) M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

14 Experimental scenarios Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation Source and target nodes of the Planet-lab experiments IP address Domain name Source S planetlab1.pop-mg.rnp.br Targets T host1.planetlab.informatik.tu-darmstadt.de T alice.cs.princeton.edu T planet1.cs.huji.ac.il T planetlab-1.cs.auckland.ac.nz T node2.planet-lab.titech.ac.jp M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

15 Experimental scenarios Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation Local testbed 10 Mbps {10, 100}Mbps 10 Mbps Mac Minis running Quagga over Ubuntu 9.04 M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

16 FLAME ping prototype Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation 1 function flameping ( params) 2 Ping params ( and corresponding defaul ts ) 3 lo c al target = params. target or lo c al si ze = params. size or 56 5 lo c al i n t e r v a l = params. i n t e r v a l or microsecond r e s ol ut i o n 6 lo c al npackets = params. npackets or 10 7 lo c al protocol = params. protocol or icmp 8 lo c al timeout = params. timeout or microsecond r e s o l u t i o n 9 lo c al t t l = params. t t l or for i = 1, npackets do 12 lo c al response Choose probing protocol ( for UDP and TCP primitives, 15 since port i s not indicated i t i s randomly chosen above 1024) 16 i f protocol == icmp then 17 response=lamp.sendicmptw{ip= target, si ze= size, 18 timeout= timeout, t t l= t t l } 19 e l s e i f proto co l == udp then 20 response=lamp.sendudptw{ip= target, size= size, 21 timeout= timeout, t t l= t t l } 22 e l s e i f proto co l == tcp then 23 response=lamp.sendtcptw{ip= target, size= size, 24 timeout= timeout, t t l= t t l } 25 else 26 print ( INVALID PROTOCOL:, protocol ) 27 return 28 end Check host/ net u nreach ab i l i t y 31 i f response and response. l oss == ICMP HOST UNREACH then 32 print ( DESTINATION UNREACHABLE :, target ) 33 return 34 end Wait time between probes 37 i f type ( i n t e r v a l ) == number then 38 lamp. sleep ( i n t e r v a l ) 39 e l s e i f type ( i n t e r v a l ) == table then 40 lamp. sleep ( i n t e r v a l. func ( i n t e r v a l. params )) end end f or i 43 end M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

17 Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation Commands to execute flameping over Planet-lab nodes Probes uniformly spaced l o ca l t argets = { , , , , } for k, v in i p a i r s ( t a rg ets ) do flameping{ target=v} end Probes exponentially spaced l oc al function expdist ( lambda ) l o c al r = 0 repeat r = math. random () un ti l ( r = 0) return math. f l o o r ( math. log ( r )/ lambda ) end l oc al t argets =... same as l e f t for k, v in i p a i r s ( t arge ts ) do flameping{ t ar get=v, i n t e r v a l={ func=expdist, params=1/ } } end M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

18 RTT measurement statistics in Planet-lab Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation ping flame ping flame ping exp RTT (ms) T1 T2 T3 T4 T5 Targets for the RTT measurements M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

19 FLAME tracing route prototype Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation 1 function flametrace ( params ) 2 lo c al target = params. target or lo c al si ze = params. size or 60 4 lo c al i n t e r v a l = params. i n t e r v a l or lo c al npackets = params. npackets or 3 6 lo c al protocol = params. protocol or udp 7 lo c al timeout = params. timeout or 5 8 lo c al maxhops = params. maxhops or for h = 1, maxhops do 11 lo c al response for i = 1, npackets do 14 Choose probing protocol ( for UDP and TCP primitives, 15 since port i s not indicated i t i s randomly chosen above 1024) 16 i f protocol == icmp then 17 response=lamp.sendicmptw{ip= target, si ze= size, 18 timeout= timeout, t t l=h} 19 e l s e i f proto co l == udp then 20 response=lamp.sendudptw{ip= target, size= size, 21 timeout= timeout, t t l=h} 22 e l s e i f proto co l == tcp then 23 response=lamp.sendtcptw{ip= target, size= size, 24 timeout= timeout, t t l=h} 25 else 26 print ( INVALID PROTOCOL:, protocol ) 27 return 28 end same as l i ne s in flameping 31 end f or i I s current node the target? 34 i f ( response. dstip == target ) then break end 35 end f or h 36 end M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

20 Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation Graph representation of traceroute and flametrace results (T1) The numbered edges indicate the number of omitted hops between the corresponding nodes (T3) (T5) (T2) (T4) M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

21 FLAME one-way delay measurement prototype Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation 1 function flameowd( params ) 2 lo c al target = params. target or lo c al si ze = params. size or 56 4 lo c al i n t e r v a l = params. i n t e r v a l or lo c al npackets = params. npackets or 5 6 lo c al protocol = params. protocol or udp 7 lo c al timeout = params. timeout or lo c al port = params. port or n i l 9 10 for i = 1, npackets do 11 lo c al response Choose probing protocol ( i f port i s not indicated the destination node 14 chooses a port above In any case, for OW operations, 15 the source and destinat i on nodes agree on the used port 16 through XMPP messages ) 17 i f protocol == udp then 18 response=lamp.sendudpow{ip= target, size= size, 19 timeout= timeout, port= port} 20 e l s e i f proto co l == tcp then 21 response=lamp.sendtcpow{ip= target, size= size, 22 timeout= timeout, port= port} 23 else 24 print ( INVALID PROTOCOL:, protocol ) 25 return 26 end Check port u n a v a i l a b i l i t y and host/ net u nreach ab i l i t y 29 i f response then 30 i f response. loss == ICMP HOST UNREACH then 31 print ( DESTINATION UNREACHABLE :, target ) return e l s e i f response. l o s s == PORT ALREADY IN USE then 34 print ( DESTINATION PORT ALREADY IN USE:, target, port ) 35 return 36 end 37 end same as l i n e s i n flameping 40 end f or i 41 end M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

22 Measured one-way delay in ms Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation Topology Topology P05 Median P95 P05 Median P95 owping flameowd NTP estimated synch error is±0.41 ms M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

23 FLAME link capacity estimation prototype Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation 1 function flamechar ( params) 2 lo c al mtu = params. mtu or 1500 probe size l im i t 3 lo c al increment = params. increment or 32 dif fe re nc e between probe si ze s 4 lo c al npackets = math. f l o o r ( mtu/ increment ) 5 lo c al maxhops = params. maxhops or 30 maximum number of allowed hops 6 lo c al r e p e t i t i o n s = params. r e p e t i t i o n s or 32 number of t e st s per hop 7 lo c al target = params. target or lo c al timeout = params. timeout or 3 9 lo c al i n t e r v a l = param table. i n t e r v a l or time between probes 10 lo c al s i z e l i s t = generateprobesizelist( npackets, increment ) for h = 1, maxhops do 13 lo c al response 14 for t = 1, r e p e t i t i o n s do 15 for i = 1, npackets do 16 local response Choose probing protocol ( for UDP and TCP primitives, 19 since port i s not indicated i t i s randomly chosen above 1024) 20 i f protocol == icmp then 21 response = lamp.sendicmptw{i p= target, s i z e= s i z e l i s t [ i ], t t l=h, timeout= timeout} 22 e l s e i f proto co l == udp then 23 response = lamp.sendudptw{i p= target, s i z e= s i z e l i s t [ i ], t t l=h, timeout= timeout} 24 e l s e i f proto co l == udp then 25 response = lamp.sendtcptw{i p= target, s i z e= s i z e l i s t [ i ], t t l=h, timeout= timeout} else print ( INVALID PROTOCOL:, protocol ) 28 return 29 end same as l i ne s in flameping 32 end f or i 33 end f or t I s current node the target? 36 i f ( response. remip == target ) then break end 37 end f or h 38 end Function to create a shuffle d package siz e l i s t 41 function generateprobesizelist( npackets, increment) implementation omitted due to space l i mi ta ti ons 43 end M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

24 Link capacity measurements Experimental scenarios RTT measurements Tracing Route One-way delay measurements Link capacity estimation RTT (ms) pchar datapoints pchar best fit (R 2 = 0.92) flamechar datapoints flamechar best fit (R 2 = 0.92) C pchar = 4.69 Mbps C flamechar = 5.03 Mbps RTT (ms) pchar datapoints pchar best fit (R 2 = 0.93) flamechar datapoints flamechar best fit (R 2 = 0.91) C pchar = 4.73 Mbps C flamechar = 5.11 Mbps Size of the probe traffic (bytes) (a) Link 1, Topology Size of the probe traffic (bytes) (b) Link 1, Topology RTT (ms) pchar datapoints pchar best fit (R 2 = 0.98) flamechar datapoints flamechar best fit (R 2 = 0.98) C pchar = 4.83 Mbps C flamechar = 4.46 Mbps RTT (ms) pchar datapoints pchar best fit (R 2 = 0.93) flamechar datapoints flamechar best fit (R 2 = 0.94) C pchar = Mbps C flamechar = Mbps Size of the probe traffic (bytes) Size of the probe traffic (bytes) (c) Link 2, Topology (d) Link 2, Topology M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

25 Conclusions Conclusions Future work Questions? FLAME provides the following contributions an environment for the rapid prototyping of active measurement tools based on a small but comprehensive set of probing primitives, allowing the implementation of tools that depend or not on cooperative destination nodes; a centralized repository for implicitly gathering measurement results in a common data format, encouraging the untangling of probing and data output functionality and easing the process of further analysis and comparison among different result datasets. M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

26 Conclusions Conclusions Future work Questions? Possible uses of FLAME prototyping besides rapid prototyping of new active measurement techniques... fair comparison among different measurement techniques easier combination of results from different tools easier tailoring of FLAME-based prototypes M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

27 Future work Conclusions Future work Questions? It s an ongoing work! Future work 1 dynamic selection of available agents 2 enrichment of the measurement toolkit with usual active measurement tools Application in other contexts porting measurement agents to resource-constrained devices taking benefit of Lua s low memory footprint M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28

28 That s all, folks! Conclusions Future work Questions? Interested? Give FLAME a try! FLAME code is available at Contact Marcos Kirszenblatt: marcoslk@lncc.br Thiago Cardozo: thiagoc@lncc.br Artur Ziviani: ziviani@lncc.br Antônio Tadeu Azevedo Gomes: atagomes@lncc.br M. L. Kirszenblatt et al. STIC-AmSud Project Meeting Dec /28