Layered Queueing Network Modeling of Software Systems

Transcription

1 1 Layered Queueing Netwrk Mdeling f Sftware Systems Building Security System (buffering) Murray Wdside 5201 Canal Building

2 Building Security System LQN-examples slide 2 Tw subsystems: CCTV strage, and dr access cntrl hpe t manage up t 100 cameras Cmpnents, shwn as UML with MARTE anntatins: SecurityCard Reader DrLck Actuatr Vide Camera Disk {PAcapacity=2} LAN <<PAhst>> ApplicCPU AccessCntrl Acces Cntrller VideAcquisitin Vide Cntrller AcquirePrc StrePrc Buffer {PAcapacity=$Nbuf} Buffer Manager <<PAhst>> DB_CPU Database

3 <<PApenLad>> {PAccurencePattern = ( pissn, 0.5, s ), PArespTime =(( req, percentile,95, (1, s )), ( pred, percentile, 95, $UserR)) } Dr Access Scenari 1. Pissn arrivals f Users 2/s <<PAcntext>> LQN-examples slide 3 User {PAextOp=(read, 1)} readcard User puts card thru reader enterbuilding CardReader DrLck admit (cardinf) Alarm mean, (1.8, ms ))} 3. Open dr r recrd alarm Requirement: 1 s respnse with 95% prbability mean, (1.8, ms ))} {PAdelay=( asmd, mean, (500, ms )), PAprb = 1} [OK] pendr() {PAprb = 0} [nt OK] alarm() mean, (1.8, ms ))} Access Cntrller Database {PAcapacity=10} mean, (0.3, ms ))} checkrights() mean, (0.2, ms ), PAprb=0.2} [need t lg?] lgevent() writeevent() Disk {PAcapacity=2} mean, (1.8, ms ))} getrights() mean, (1.5, ms )), PAprb = 0.4} readrights() [nt_in_cache] readdata() 2. Prcess card, check rights mean, (3, ms ))} writerec() 4. Write lg f events

4 CCTV capture scenari LQN-examples slide 4 <<PAcntext>> Vide Cntrller 1.Trigger camera read events (the N cameras were nt mdeled) 2. Put the image in a buffer, send t database *[$N] prconeimage(i) mean, (1.8, ms))} {PArep = $N} <<PAclsedLad>> {PAppulatin = 1, PAinterval =(( req, percentile,95, (1, s )), ( pred, percentile, 95, $Cycle)) } AcquirePrc {PAdemand =( asmd, mean, (1.5, ms )} getbuffer() <<GRMacquire>> allcbuf (b) mean, ($P * 1.5, ms )), PAextOp = (netwrk, $P)} getimage (i, b) mean, (0.9, ms ))} passimage (i, b) mean, (0.5, ms ))} BufferManager This bject manages the resurce Buffer mean, (0.5, ms ))} mean, (1.1, ms ))} streimage (i, b) <<GRMrelease>> releasebuf (b) StrePrc mean, (0.2, ms ))} freebuf (b) mean, (2, ms ))} stre (i, b) Database {PAcapacity=10} Buffer peratins: get release mean, ($B * 0.9, ms )),, PAextOp=(writeBlck, $B)} writeimg (i, b) 3. Stre image in database

5 The LQN Mdel LQN-examples slide 5 Shared Resurces Vide Capture subsystem Dr Access subsystem

6 Handling f Buffering Each buffer must be emptied befre it can be used fr anther camera Thus the buffer is a resurce that culd have a queue, which shuld be mdeled as the pseud-task Buffer Mdel fragment withut buffer Hw the buffer pl was mdeled LQN-examples slide 6 real task BufferManager and pseudtask Buffer frwarding call t a functin f BufferManager

7 Handling f Buffering (2) LQN-examples slide 7 The peratins that require hlding the buffer are executed by calls frm the buffer pl pseud-task Buffer separate frm the buffer manager task!! including the executin f the release peratin by the buffer manager assumes the manager has a dedicated thread fr release releasebuf is executed by streimage, r bufentry

8 Mdel Features LQN-examples slide 8 This mdel illustrates hw we can mdel lgical resurces (the buffer pl) the use f frwarding the use f secnd phase t imprve cncurrency (later)

9 Results #1 LQN-examples slide 9 Average Nrmalized Utilizatins Prb f Missing Deadline Respnse Time Ncam Cycle User (sec) AcqPrc Buffer StrePrc AppCPU Cameras Drs (sec) Table 1. Simulatin results fr the base case Base case, ne buffer, s ne camera at a time Access-cntrl respnses are fine; the event rate was kept cnstant at 2/s. Camera plling becmes t slw between 20 and 30 cameras try adding mre buffers.

10 Results #2 cameras fixed at 40, vary the number f buffers NBuf LQN-examples slide 10 disappinting: the miss prbability levels ut abve 9% at abut 7 buffers. StrePrc is apparently the new bttleneck: try an additinal thread Average Prb f Missing Nrmalized Utilizatins Respnse Time Deadline NBuf Cycle User AcqPrc Buffer StrePrc AppCPU Cam s Drs (sec) (sec)

11 Results #3 LQN-examples slide 11 Success! Tw threads n StrePrc cmbined with 4 buffers brings the miss prbability dwn well within spec f 1 secnd fr each N. f Stre Prc Average Respnse Time Cycle (sec) 40 cameras, Nuser = 100 drs, Nbuf = 4 buffers User (sec) Nrmalized Utilizatins Prb f Missing Deadline AcqPrc Buffer StrePrc AppCPU Cam s Drs

12 Discussin LQN-examples slide 12 The saturated resurce is AppPrc making multiplicity = 2 allws 50 cameras The limitatin is nw at AcquirePrc, due t a lng service time the time it takes t stre the buffers is limiting multithreading alne is nt the answer T allw an earlier start n the next camera, we can put the calls frm AcquirePrc int phase 2, with multithreading early reply t VideCntrller mves the capture n t the next camera much earlier allws the cncurrent phase-2 Acquire tasks t run in parallel Other adjustments are als pssible

13 Results #4 LQN-examples slide 13 By using phase 2 at AcquirePrc and varius multiplicities we can get a capacity f 100 cameras. Even mre capacity can be fund with StrePrc. Anther explratin apprach: set multiplicities at inf and see if specified delays are feasible at all, and what mulitplicity is used (= utilizattin), then wrk dwn t specified delays. Multiplicity (Acquire, Buffer, Stre, App. CPU) 100 cameras, Nuser = 100 drs, Nbuf = 10 buffers Average Respnse Time Cycle (sec) User (ms) Acquire Prc Nrmalized Utilizatins Stre Prc Prb f Missing Deadline Buffer App Cam s Drs CPU 2, 4, 2, , 10, 6, , 10, 6,