Soft Real Time Design

Transcription

1 by Gerrit Muller HSN-NISE Abstract Soft Real Time design addresses the performance aspects of the system design, under the assumption that the hard real time design is already well-covered. Core decisions in soft real time design are: granularity synchronization prioritization allocation resource management The complete course ASP TM is owned by TNO-ESI. To teach this course a license from TNO- ESI is required. This material is preliminary course material. status: preliminary draft

2 hard real time soft real time disastrous failure failure dissatisfaction irritation human safety device safety loss of information loss of functionality or (image) quality loss of eye hand coordination limited throughput waiting time 2 Gerrit Muller PSRTpositioning

3 Case 1 TV zapping Problem introduction Approach for solving response time problems Revised functional model Measuring and modelling 3 Gerrit Muller PSRTzapIntro

4 Zap timing: What is the Requirement? new channel P+ P- zap total response time zap repetition remote control zap visual feedback open for next respons new channel visual feedback time time 4 Gerrit Muller EACresponseTime

5 Approach 1) Measure the end-to-end time 2) Decompose the processes based on expected outcome 3) Measure the individual components use previous decomposition (2) 4) Clarify the unknown parts and make them explicit 5) Further divide the major posts 6) Aggregate the smaller posts 5 Gerrit Muller PSRToptimizeTimingIssues

6 Functional Model User Interface ~100 ms Line demux: ~ 60µs Video signal de-mux Bit detection ~ 150 ns Teletext processing User i/f graphics generation Teletext overlay generation Video signal mux Blank Tuner de-mux Picture processing User Interface ~100 ms Audio processing Audio / video sync ~ 20ms Mute Control User Interface ~100 ms ~1.8ms / bit 6 Gerrit Muller PSRTzappingFunctionalModel

7 Expectated and Measured Values Expected values: Measured values: Mute Blank Flush AV pipeline Set tuner Fill AV pipeline Unmute Unblank : 50 ms : 40 ms : 160 ms : 200 ms : 160 ms 1 frame 4 frames : 50 ms : 40 ms 4 frames Mute : 60 ms Blank : 120 ms Flush AV pipeline : 0 ms Set tuner : 180 ms Fill AV pipeline : 40 ms Format detection : 200 ms Unmute : 60 ms Unblank : 120 ms Summing : ~ 900 ms Total time measured: 2000 ms 1 frame 5 frames 7 Gerrit Muller PSRTzappingProblemStep3

8 Analysis and Improvements Zapping Problem step 4 Somewhere 1000 ms are missing Detection of frame size takes a long time! + Lots of software overhead Analyze frame size detection and SW overhead Zapping Problem step 5 Subdivide / analyze format detection (200 ms) Zapping Problem step 6 Ignore pipeline effects 8 Gerrit Muller PSRTzappingProblemAnalysis

9 Simple Concurrency Model (with waits) Zapping tasks sequential Video present Set Tuner No video Blank Video Detect Framesize Video present Flush AV pipeline Fill AV pipeline Mute Audio Zap Blink LED OSD Zap finished 9 Gerrit Muller PSRTzappingSequential

10 Simple Concurrency Model (optimized) Zapping tasks parallel Set Tuner Video No video Blank Video Detect Framesize Video present Flush AV pipeline Fill AV pipeline Mute Audio Zap Blink LED OSD Zap finished 10 Gerrit Muller PSRTzappingConcurrent

11 Case 1 Summary TV zapping Understanding of the problem is crucial Iterate over modelling and measuring to build balanced performance model 11 Gerrit Muller PSRTzapSummary

12 Case 2 EasyVision: Resource Management Introduction to application SW design Memory and performance Memory design CPU load and Performance 12 Gerrit Muller PSRTeasyvisionCaseIntro

13 Introduction to Medical Imaging Application Easyvision Medical Imaging Workstation serving 3 X-ray examination rooms providing interactive viewing and printing on high resolution film Challenge: interoperability and WYSIWYG over different products 13 Gerrit Muller PSRTeasyvisionIntro

14 Easyvision Serving Three URF Examination Rooms typical clinical image (intestines) URF-systems EasyVision: Medical Imaging Workstation 14 Gerrit Muller MSeasyVision

15 Image Quality Expectation WYSIWYG image generation X-ray system presentation 3 rd party workstation monitor film network, storage monitor what you see at one work-spot is application processing Easyvision presentation monitor film network, storage what you get at another work-spot??? 15 Gerrit Muller MICVwysiwyg

16 Presentation Pipeline for X-ray Images image from database interpolate Look up table invert contrast / brightness graphics merge colour LUT monitor spatial enhancement bi-linear bi-cubic output brightness contrast legend input SW HW 16 Gerrit Muller MICVpresentationPipeline

17 Quadruple View-port Screen Layout UI icons, text view-port 1 view-port pixels ca 200 pixels viewport 5 view-port 3 view-port pixels ca. 460 pixels 17 Gerrit Muller MICVquadrupleViewportLayout

18 Rendered Images at Different Destinations Screen: low resolution fast response Film: high resolution high throughput Network: medium resolution high throughput 18 Gerrit Muller MICVdestinations

19 SW Design Easyvision SW design Concurrency design SW layers 19 Gerrit Muller PSRTeasyvisionSWdesignIntro

20 Concurrency via Software Processes remote systems and users user client legend user control communication user interface client process export data base optical storage UI devices print system monitor Unix daemons server process control and data flow operational process network disk drive optical disk drive printer associated hardware 20 Gerrit Muller MICVsoftwareProcess

21 Criteria for Process Decomposition management of concurrency management of shared devices unit of memory budget (easy measurement) enables distribution over multiple processors Processes are a facility provided by the Operating System (OS) to manage concurrency, resources and exceptions unit of exception handling: fault containment and watchdog monitor 21 Gerrit Muller PSRTprocessCriteria

22 Simplified Layering of the SW (Construction Decomposition) dev. tools service SW keys Config Install Start up Spool HCU Store Image Gfx UI DB RC driver RC interf RC HC driver HC interf 3M DOR driver DOR Medical Imaging R/F Print Store View Cluster SunOS NIX Standard IPX workstation Desk, cabinets, cables, etc. DSI PMSnet in PMSnet out legend user interface application functions toolbox operating system hardware SW infrastructure connected system 22 Gerrit Muller MICVswLayers1992

23 Memory Use and Performance Easyvision Memory and Performance Performance problems Analysis of memory use Memory budget 23 Gerrit Muller PSRTeasyvisionMemoryIntro

24 Performance as a Function of Memory Use performance Good Bad 0 memory usage physical paging to disk memory 64 MB 200 MB 24 Gerrit Muller EASRTperformanceVsMemory

25 Problem: Unlimited Memory Consumption (1992) total measured memory usage OS code data bulk data fragmentation performance 0 MB memory usage physical paging to disk memory MB 25 Gerrit Muller MSmemoryZeroMeasurement

26 Measurement Per Process data MByte UNIX shared libraries UI communication server storage server print server other code measured (left column) budget per process (right column) 26 Gerrit Muller MSmemoryBudget

27 Solution: Measure and Iterative Redesign 200 MB fragmentation anti-fragmenting bulk data data code budget based awareness, measurement DLLs 74 MB OS measured tuning budget 27 Gerrit Muller MSmemoryUsageReduction

28 Method: Budget per Process Budget: + measurable + fine enough to provide direction MByte dll's UI communication server storage server print server other UNIX measured (left column) + coarse enough to budget per process (right column) be maintainable 28 Gerrit Muller MSmemoryBudgetAnnotated

29 Example of a Memory Budget memory budget in Mbytes code obj data bulk data total shared code User Interface process database server print server optical storage server communication server UNIX commands compute server system monitor application SW total UNIX Solaris 2.x file cache total Gerrit Muller RVmemoryBudgetTable

30 Exercise: Bulk Data Capacity Memory block 12MByte How many blocks of 1024 x bits data can be stored? How many blocks of 1024 x bits data can be stored? 30 Gerrit Muller PSRTexerciseBulkdata

31 Exercise: Object Data Capacity Object Data 3MByte Frequency Description Typical size 1 Large objects (e.g. dictionary) 20 kb 20 Medium object, e.g. UI data 200 Bytes 1000 Small object, e.g. image attributes 20 Bytes Total How many objects with this distribution fit in the 3MByte Object data store? 31 Gerrit Muller PSRTexerciseObjectData

32 Memory Budget of Easyvision RF R1 and R2 code object data bulk data total memory budget in Mbytes R1 R2 R1 R2 R1 R2 R1 R2 shared code UI process database server print server DOR server communication server UNIX commands compute server system monitor application total UNIX file cache total Gerrit Muller MICVmemoryBudgetR1R2

33 Answer: Bulk Data Capacity Memory block 12MByte How many blocks of 1024 x bits data can be stored? How many blocks of 1024 x bits data can be stored? 12 6 Memory block 12MByte Memory block 12MByte * Assuming that 8-bit data is stored as 8-bit (char) Assuming that 16-bit data is stored as 16-bit (short int) 33 Gerrit Muller PSRTanswerBulkdata

34 Answer: Object Data Capacity Object Data 3MByte Frequency Description Typical size Size * Freq 1 Large objects (e.g. dictionary) 20 kb 20 kb 20 Medium object, e.g. UI data 200 Bytes 4kB 1000 Small object, e.g. image attributes 20 Bytes 20kB Total 44kB 44kByte fits approximately 68 times in 3MByte Expect to store at most 68 large objects (1360 Medium sized objects, small objects) 34 Gerrit Muller PSRTanswerObjectData

35 Memory Use and Performance Easyvision Memory Design Fragmentation and consequences Application caches Memory design applied 35 Gerrit Muller PSRTeasyvisionMemoryDesignIntro

36 Memory Fragmentation image 1, 256 kb image 2, 256 kb image 3, 256 kb 1. replace image 3 by image 4 image 1, 256 kb image 3, 256 kb legend image in use image 1, 256 kb 4 2. add image 5 image 1, 256 kb 4 3. replace image 1 by image 6 4 image 3, 256 kb image 3, 256 kb image 3, 256 kb unused memory image 5, 256 kb image 5, 256 kb 6 4 image 3, 256 kb image 5, 256 kb 36 Gerrit Muller MICVfragmentationAnimation

37 Memory Fragmentation Increase MBytes used address space gross used nett used time 37 Gerrit Muller MICVfragmentationInTime

38 Cache Layers Medical imaging R/F cache sizes legend cluster PixMap cache print PixMap cache allocator, chunk view PixMap cache user interface application functions heap memory, malloc() free() toolbox memory management unit virtual memory instruction and data cache physical memory disk storage operating system hardware 38 Gerrit Muller MICVcacheLayers

39 Bulk Data Memory Management Memory Allocators chunk size:3mb chunk size: 1MB for stamp images 96*96*8 (9kB) block size: 9kB for large images from 225 kb (480*480*8) to 3 MB (1536*1024*16 ) block size: 256kB chunk size: 2MB for small (screen) images from 8kB to 225 kb block size: 8 kb 39 Gerrit Muller MICVmemoryAllocators

40 Cached Intermediate Processing Results text gfx raw enhanced retrieve image enhance image interpolate resized image greyvalue viewport lookup image merge display 40 Gerrit Muller MICVprocessingCachedPixmaps

41 Example of Allocator and Cache Use bit image requires 4 256kB blocks images require kB blocks 12 blocks shortage block size: 256kB block size: 9kB block size: 8 kb image 8 bit requires 27 8kB blocks images require 5 8kb blocks all screen-size images require 334 8kB blocks, 78 blocks shortage text Pixmap cache gfx greyvalue viewport viewport raw raw resized greyvalue viewport viewport raw raw resized raw image raw resized image greyvalue image resized raw image enhanced image image resized image image image image image image image retrieve enhance image interpolate lookup image merge display 4 * byte / pixel 4 * byte / pixel 4 * byte / pixel 4 * byte / pixel 4 * byte / pixel retrieve enhance interpolate lookup merge display 41 Gerrit Muller MICVpixmapExample

42 Print Server is Based on Banding 1024 pixels 1024 pixels original images 128 pixels 4k pixels 42 Gerrit Muller MICVbanding

43 CPU Load and Performance Easyvision Memory CPU load and performance CPU load analysis response time throughput measurement tools 43 Gerrit Muller PSRTeasyvisionCPUloadIntro

44 CPU Processing Times and Viewing Responsiveness txt gfx retrieve 0.3s raw image enhance 0.5s enhanced image viewport interpolate 0.2s resized image lookup (LUT) 0.075s display greyvalue image gfx merge 0.025s display 0.05s next update rate for zoom C/B 0.9s -1 common user actions 3 s -1 7 s -1 retrieve enhance interpolate LUT g f x accumulated processing time in seconds pipeline timing proportional 44 Gerrit Muller MICVprocessingTimes

45 Server CPU Load remote systems and users communication 2.5 CPU second per Mbyte input import serving one examination room CPU time available for interactive viewing serving 3 examination rooms margin 2 min import 2.5 min / exam data base print printer disk 3.5 CPU second per Mpixel output print 50 s/exam 210 s/exam 30% print 10.5 min / exam 90% 45 Gerrit Muller MICVserverCPUload

46 Resource Measurement Tools t n-2 t n-1 t n preamble to remove use case start-up effects time oit ps vmstat kernel resource stats object instantations heap memory usage kernel CPU time user CPU time code memory virtual memory paging heapviewer (visualise fragmentation) 46 Gerrit Muller MICVtools

47 Object Instantiation Tracing class name current nr of objects deleted since t n-1 created since t n-1 heap memory usage AsynchronousIO AttributeEntry BitMap BoundedFloatingPoint BoundedInteger BtreeNode1 BulkData ButtonGadget ButtonStack ByteArray [819200] [ ] [13252] 47 Gerrit Muller MICVoitTool

48 Overview of Benchmarks and Other Measurement Tools test / benchmark what, why accuracy when self made public SpecInt (by suppliers) CPU integer coarse new hardware Byte benchmark computer platform performance OS, shell, file I/O coarse new hardware new OS release file I/O file I/O throughput medium new hardware image processing CPU, cache, memory as function of image, pixel size Objective-C overhead method call overhead memory overhead socket, network data base throughput CPU overhead transaction overhead query behaviour accurate accurate accurate accurate new hardware initial ad hoc ad hoc load test throughput, CPU, memory accurate regression 48 Gerrit Muller MICVbenchmarks

49 Case 4 MRI Volume Reconstruction and Viewing Usage patterns as impact on performance Resource model and requirements identification for usage patterns 49 Gerrit Muller PHRTcaseMRvolumeIntro

50 Volume Acquisition and Reconstruction 256 Data in bytes = 2 * 512 * 512 * 256 * 2 = 512 Volumes x y z bytes per pixel 256 MBytes 512 in 2 * 2 minutes = 240 seconds 50 Gerrit Muller MRneuroCubic

51 Performance Requirements Examination of previous patient 15 minute time slot George arrives at radiology department Nurse explains the procedure George is waiting in the dressing room Prepare George for the examination George leaves exam room (a.o. RF coils) Position Imaging View away View away 14:00 14:15 14:30 51 Gerrit Muller MRneuroTypicalTimeline

52 Resource Model Acquisition Reconstruction Intermediate data: 256 MByte Viewing View away in ca 10 sec. Storage full screen 25 images per second 2 Volumes 256 MByte 52 Gerrit Muller MRneuroResourceModel

53 Critical Resources Buffer architecture Pipeline & caching Acquisition Reconstruction Intermediate data: 256 MByte Viewing View away in ca 10 sec. Attribute access Storage 2 Volumes 256 MByte full screen 25 images per second 53 Gerrit Muller MRneuroResourceCriticalities

54 Case 4 MRI Volume Reconstruction and Viewing Operational usage pattern drives (implicit/explicit) system performance requirements Resource / cost trade-off must support operational usage patterns 54 Gerrit Muller PHRTcaseMRvolumeSummary

55 Case 5 Mobile Display Appliances Modelling external environment End-to-end performance Allocation choices 55 Gerrit Muller PSRTmobileApplianceIntro

56 Mobile Display Appliances Mobile Display Appliance Mediascreen Original pictures from Nokia 56 Gerrit Muller FFTSclient

57 User Access Point to a Long Foodchain Home Server Network Providers Service Providers Content Providers Appliance User 57 Gerrit Muller FFTStotalChain

58 The SMART World of the Design Standard Interactive System Applications Data transport Security Virtual Machine Display and UI free after Nick Thorne, Philips Semiconductors, Systems Laboratory Southampton UK, as presented at PSAVAT April Gerrit Muller FFTSstandardInteractiveSystem

59 Specifiable Characteristics Servers & Networks Standard Interactive System Throughput Latency Distance Power Data transport Security Level Performance Encryption Authentication Security Functionality Performance Power, Footprint Applications Virtual Machine Functionality Performance Power, Footprint Display size Color depth Rendering Performance IQ UI modi Display and UI 59 Gerrit Muller FFTSstandardInteractiveSystemAnnotated

60 Response Time: Latency Budget times in milliseconds Appliance Data transport Security Virtual Machine Application Graphics and UI Home Network Home Server Network contention Provider Infrastructure Total Last-Mile network Backbone network Service server Content server User need Message Latency Response Time All numbers are imaginary and for illustration purposes only 60 Gerrit Muller FFTSlatencyBudget

61 Interaction or Irritation? Response Time (ms) Interactive Experience Irritating Experience Home Server Network Providers Service Providers Content Providers Appliance User 61 Gerrit Muller FFTStotalChainResponseTime

62 Case 5 Summary Mobile Display Appliances Modelling external environment: make assumptions End-to-end performance: large part of performance budget is not controlled User perceived performance determines function allocation 62 Gerrit Muller PSRTmobileApplianceSummary

63 Colofon The ASP TM course is partially derived from the EXARCH course developed at Philips CTT by Ton Kostelijk and Gerrit Muller. Extensions and additional slides have been developed at ESI by Teun Hendriks, Roland Mathijssen and Gerrit Muller. 63 Gerrit Muller PERFcolofon