Challenges of Flexible Real-Time Communication

Transcription

1 -1 - ARTIST2 Summer School 2008 n Europe Autrans (near Grenoble), France September 8-12, Challenges of Flexble Real-Tme Communcaton Lecturer: Lus Almeda Embedded Systems Lab IEETA Unversty of Avero, Portugal

2 -2 - Flexblty? Flexblty the capablty of changng form a ready capablty to adapt to new, dfferent, or changng requrements (Merram-Webster on-lne) Wthn embedded systems, flexblty s a vast property that can apply to off-lne or on-lne features a desgn flow that supports changes wth mnor redesgn effort systems prepared for dfferent confguratons at/after deployment hardware that supports mult-functonalty systems that adapt to varyng run-tme condtons (load, users, resources...) systems that reconfgure upon hazardous events or occasonal operaton of certan components...

3 -3 - Flexblty ncreasng nterest Appears frequently as Adaptablty and Reconfgurablty Topc of an actvty wthn ArtstDesgn (NoE n FP7) Other European projects (DySCAS,FRESCOR,RUNES,HPEAC...) Several related events NeRES 2007, APRES 2008, ARC ,... EU FP7 Embedded Systems Call The engneerng of more robust, context-aware and easy-to-use ICT systems that self mprove and self-adapt wthn ther respectve envronments. Heterogenety; composablty; predctablty of extra-functonal propertes; adaptvty for copng wth uncertanty; and unfcaton of approaches from computer scence, electronc engneerng and control.

4 -4 - In ths talk Focus on Operatonal Flexblty (on-lne) n Dstrbuted (networked) Embedded Systems What we want Two possble approaches Some challenges Our recent research nlne wth these approaches Concluson & general ssues

5 -5 - What we want To be able to connect any component to the system, on-lne, beng sure that: Nothng bad wll happen The system wll do ts best to ntegrate the new component: It can accept the new component wthout any adjustment It can accept t upon system adjustment It can reject the new component Reconfgurablty

6 -6 - What we want Allow the system to adjust on-lne accordng to effectve nstantaneous needs or resource avalablty: Free and reuse the resources of subsystems that operate occasonally or fal when off Adapt resources requred by each subsystem on-lne to: Mnmze the resources used (e.g. BW, energy, ) Maxmze the servce delvered wth a fxed level of resources Adaptvty Cope wth varyng levels of resource avalablty (e.g. lnks wth varable BW, decreasng avalable energy, )

7 -7 - What we want Focus on Dstrbuted Embedded Systems: Work on Flexble schedulng, Feedback schedulng, Elastc schedulng, Resource usage optmzaton n control systems,..., has manly addressed unprocessors There have been few extensons to dstrbuted embedded systems In these, the network s central and must support the desred level of flexblty

8 -8 - Two approaches to acheve flexblty n dstrbuted (networked) embedded systems Law enforcement Negotate resource needs and enforce ther proper use Requres good level of control over resources Hard real-tme Admsson controller, QoS manager,... Voluntary cooperaton Assume nodes wll cooperate for proper resource usage No enforcement unavodable lmtatons caused by non-cooperatng nodes, nterference,... Best effort

9 -9 - Law enforcement Some challenges How strong/robust s the enforcng of proper resource usage? Control over resources and flexblty management mply extra resource needs (BW, CPU, energy...)! Also mply extra complexty! Wth potental for lower relablty! Extra state nformaton s needed to manage flexblty. Is t replcated? If so, t must be consstent... Is there a safe state n case of nconsstency? Voluntary cooperaton Some knd of guarantees (probablstc) would be welcome! Characterzaton of operatonal envronments Varyng communcaton lnks Resource usage optmzatons

10 -10 - Some challenges Flexblty management How to dstrbute free BW among a set of users? Elastc models (m,k)-frm model Greedy models... Act on C, on T, on both... When and how to trgger adaptaton / reconfguratons? Flexble mode changes... How to relate resource usage and QoS? Or even better, QoE?

11 -11 - Our recent research FTT framework new developments FTT-SE (enhanced aperodc communcaton) Dynamc QoS management (dstrbuted vdeo system) Server-SE (server-based communcaton management) Enhanced swtch (boostng robustness and ntegraton capab.) Wreless communcaton for teams of robots Adaptve-TDMA framework Round reconfguraton Flexble Tme-Trggered archtecture Law enforcement Voluntary cooperaton

12 -12 - FTT-SE Flexble Tme-Trggered communcaton over Swtched Ethernet Keepng under control the traffc load submtted to a swtched network Schedule traffc per cycles Submt only the traffc that ft n a cycle Elmnate memory overloads FTT master Support full prorty schedulng Trgger message SRT A(...) B(...)... Broadcast to all nodes l u j Nodes reacton to the TM n a gven EC l d j Ethernet swtch TM sched SM FP, EDF,... TM ε FTT master M nodes Swtch wth M ports

13 -13 - sched FTT-SE Schedulng model for perodc traffc TM Set of perodc streams (synchronous traffc) SRT = {SM : SM (C, D, T, O, Pr, S, {R 1.. R k }), =1..N s } Schedulng wth multple queues Strctly confned to the Synchronous Wndow per EC TM tr Global ready queue Sync wndow LSW l u j l d j EC... Async wn... Uplnks / Downlnks ε ε tme LSW LSW Schedulng equaton max j max j SM l C SM l LSW Memory bounds u j max ) max j=1..m (μ n j, μ p j) < (LSW e)*r/8 d j ( f max SM LSW l u j ε

14 -14 - FTT-SE Testng schedulablty of perodc traffc Basc schedulng model: Schedule wthn parttons wth strct tme bounds Use nserted dle-tme (X) There s no blockng Any analyss for preemptve schedulng can be used wth nflated transmsson tmes (C ) X n EC n EC n+1 Tme duraton E X n - Inserted dle-tme network varables We wll consder C as C and use preemptve analyss n the followng C ' = C * E E X max Inserted dle-tme compensaton factor

15 -15 - FTT-SE Testng schedulablty of perodc traffc Interference n the uplnks appears at the downlnks as release jtter (J) Utlzaton bounds are mportant for on-lne QoS management Wth release jtter they can be appled to each lnk separately TM Master u d u Node A d u Node B d u Node C d TM ε TM TM tr Synchronous Wndow LSW SM1 SM2 SM3 SM4 SM5 SM6 SM2 SM4 SM3 SM6 max J C j j j= 1.. lub 1.. n + U RM, EDF ( ) T T = j= 1 For each lnk: n = 1 C T j max 1.. n + = T 1 J U lub RM, EDF ( n) f 6 tme

16 -16 - Aperodc traffc FTT-SE Set of sporadc streams (asynchronous traffc) ART = {AM : AM (C, D, mt, Pr, S, {R 1.. R k }), =1..N a } Scheduled after the synchronous traffc Non-real-tme traffc (IP ), scheduled after the async one δ AM arrves Lsg Sched. Sgnalng message AM polled AM transmtted Lpoll TM AM_Rt = Lsg + Lsch + Lpoll

17 -17 - FTT-SE Same trggerng for all traffc Aperodc traffc s sgnaled to the Master All traffc scheduled n an ntegrated way Synchronous + asynchronous RT + Non-RT Everythng encoded n the TM

18 -18 - All aperodc Server-SE Uses aperodc mechansm of FTT-SE All traffc handled through servers Servers controls encoded n the TM Ams at managng servers dynamcally Stream wth randomly arrvng requests Same stream after passng through a sporadc server wth one packet every 5 ECs

19 -19 - Server-SE Total load vared between 18Mbt/s and 91Mbt/s 1 short control packet together wth 2 long vdeo frames

20 -20 - QoS adaptaton wth FTT-SE FTT master Two dmensonal problem: TM Adaptng a VBR source to a CBR channel Adaptng the CBR channels bandwdth explot BW released by streams that are off reduce the use of too strong compresson upon operator request CBR channels MJPEG encoders VBR bt streams

21 -21 - A dstrbuted montorng system VBR CBR adaptaton q s the compresson parameter It determnes the sze of each frame Typcal model of stream BW (R) and q R ( q) = α + β q λ q k + 1 k + 1 = k + 1 β R α (1/ λ ) β k + 1 k k = Δ + 1 R, ( k ) λ k q + β

22 -22 - A dstrbuted montorng system Adaptng multple CBR channels Streams are not always ON Maxmze total BW usage

23 -23 - A dstrbuted montorng system Adaptng multple CBR channels Evoluton of the Qualty Index (QI) comparng to statcally allocated channels QI = σ σ f fg σ g 2 fg ˆˆ ˆ 2 f + gˆ 2 2σ σ σ f 2 f σ g 2 g

24 -24-5 Mbt/s 1 Mbt/s V after 20s V 1 Mbt/s 5 Mbt/s

25 -25 - Internet browser Resource-reservaton swtch Provdng tmelness, flexblty and hgh robustness n swtched Ethernet networks Enforce negotated channel characterstcs (polcng) Reject abusve negotated traffc (flterng) Confne non-negotated traffc to separate wndows (selecton) Internet RR-swtch RR-swtch RR-swtch

26 -26 - Resource-reservaton swtch Aperodc traffc confnement TM tme EC (1ms) RT Wndow NRT Wndow Submtted traffc 1000B packets, T avg =250μs Regularty of the TM T_TM avg = 1,000ms T_TM max = 1,0003ms T_TM mn = 0,99998ms STD_TM = 138ns Outgong traffc Offset wrt the prevous TM NRT wndow - 50% EC

27 -27 - Robust IEEE communcaton Provdng robust IEEE communcaton for teams of autonomous moble robots Real-tme s best-effort Open medum, uncontrolled envronment / load, non-statonary nterference... Adaptve technques can help reducng chances of packet losses

28 -28 - Robust IEEE communcaton Adaptve TDMA Maxmzes separaton between transmssons n the team Synchronzes on receptons (no need for clock sync) Shfts phase of TDMA round to match perodc nterference Tme constrants round perod T tup Fully dstrbuted Non-synchronzed perodc broadcast Inter-packet delay (ms) M 1, M 2, M 3, M 4, M 1, T xwn Inter-packet delay (ms) Tme (s) Adaptve TDMA (less losses) M 1, t now T xwn T tup M 2, M 3, δ M M 1, 4, 4 δ 4 +1 t next Tme (s)

29 -29-2 runnng robots: Robust IEEE communcaton Dynamc reconfguraton of the slot structure Robots jon and leave dynamcally crash, mantenance, movements... Slot structure of TDMA round does not need to be predefned Number of slots contnuously adjusted to number of robots Fully dstrbuted mnmal a pror knowledge T tup δ 4 M 1, δ 4 M 4, M 1,+ 1 t now 3 runnng robots: T xwn T tup δ 4 t next M 2, M 1, M 4, M 1,+ δ 4 1 t now T xwn t next

30 -30 - Robust IEEE communcaton only 1 robot 2nd robot jons 2nd leaves 3rd jons

31 -31 - Concluson Adaptng / reconfgurng a dstrbuted system on-lne requres approprate support from the network Two man approaches can be followed: Law enforcement Such as the FTT framework, amng at controlled adaptaton/reconfg. (tmelness guarantees, traffc solaton ) Voluntary cooperaton Wreless communcaton for teams of autonomous agents, usng adaptve technques can help reducng packet losses n the presence of uncontrolled load (best effort way) A few general ssues How much control have we got over system resources? Whch guarantees do we need? How much does flexblty cost? How much do we gan?

32 -32 - Announcement CberMouse@RTSS Control the team of 5 robots wth ad-hoc communcaton capabltes to reach the vctm n the least tme A students desgn competton vctm robot Start area