Introduction to Real Time Java

Editor's Notes

• #6: The first misconception is that real-time is minimal application latency and/or maximum data throughput . But actually real-time is about derminism. A system with latencies in the range of hundereds of milliseconds can also be real-time. Nor is real-time solely for embedded applications and systems; there is a growing need for enterprise determinism. For example financial applications and telecommunication applications have very strict real-time requirements. Determinism and throughput are usually inversely related. Determinism usually cost performance. A system that needs to bring its base-case and worst-case performances as close together as possible cannot use hints or heuristics and cannot rely on the ”80/20” rule. For example a quicksort ca takeO(n2) time so although a mergesort is slower than quicksort on average it is more suitable to real-time systems because it is predictable. The resulting software is typically slower than a software that is designed to optimize typical performance, but its worst-case performance may be an order of magnitude better than such a conventional design.
• #7: Saying a system is real-time does not mean that all parts of that system is real-time. It means there are certain tasks that should be performed in real-time.
• #8: In a hard real-time system, a late answer is the same with no answer or wrong answer. In a soft real-time system, a late answer is still useable but value of it decreases rapidly. In other words occasional dead-line misses are acceptable. The severity of the consequence of missing a deadline has nothing to do with the definition of hard versus soft.
• #9: As the requirement for deterministic behavior moves from tightly coupled embedded tasks that control or monitor hardware subsystems to highly complex enterprise applications, software complexity and deterministic latency requirements become inversely proportional to solution needs and capabilities. General real-time software usually deals with time measured in milliseconds. Most real-time systems fall into this range. These systems can be programmed with normal tools. In practice, achieving response times below tens of microseconds requires a combination of custom hardware and software, possibly with no -- or a very thin -- operating-system layer. Within this overall complexity, architects, systems designers and developers need an environment that can span a significant portion of the response spectrum. This model must not only meet a broad bandwidth of performance and latency requirements, but also be scalable and relatively simple to use.
• #10: Aim of a real-time system is to put a deterministic bound to latency not to minimize it. The goal is to make latency known, small enough, consistent and measureable quantity. Jitter is the variation in latency.
• #13: As you can see worst case execution time can be 1 million times of best execution time. There can be other factors not included in this example. For example SMI (System Management Interrupts) on X86 hardware may add extra delay.
• #14: Even non rt windows and linux allows memory of a process to be locked in to memory. So at worst everything your program needs will be in memory. Some DMA controllers can be tuned to use no more than a specific faction of memory bandwidth or get off the bus entirely when the processor is servicing interrupts TLB is limited resource on processors. They are used to translate virtual addresses to physical addresses. Normally most operating systems uses 4K pages. For each page a TLB entry is required. You can reserve some number of large pages like 256Megabytes for your process.
• #15: Pinning Your Threads to processors can help to increase determinism. Both solaris and linux supports processor pinning. Solaris also supports interrupt shielding. Not all linux distributions support ineterrupt shielding. Your accound should have sufficient privileges to use processor pinning. As you can see you you can not select which thread works on which CPU. Your RealTime and NoHeapRealTime threads will use assigned cpus and all other applications and other JVM Threads (GC threads, JIT compiler thread etc.) will use other cpus in the system.
• #16: Beginning with Java SE 5.0 there is a cross-platform flag for requesting large memory pages:  -XX:+UseLargePages  (on by default for Solaris, off by default for Windows and Linux). The goal of large page support is to optimize processor Translation-Lookaside Buffers. A Translation-Lookaside Buffer (TLB) is a page translation cache that holds the most-recently used virtual-to-physical address translations. TLB is a scarce system resource. A TLB miss can be costly as the processor must then read from the hierarchical page table, which may require multiple memory accesses. By using bigger page size, a single TLB entry can represent larger memory range. There will be less pressure on TLB and memory-intensive applications may have better performance. Using large memory pages may negatively affect performance of system if it cause memory shortage for the rest of the system.
• #18: A standard Linux kernel provides soft RT behavior, and although there's no guaranteed upper bound on how long a higher-priority thread waits to preempt a lower-priority thread, the time can be roughly approximated as tens of milliseconds In RT Linux, almost every kernel activity is made preemptible, thereby reducing the time required for a lower-priority thread to be preempted and allow a higher-priority one to run. Remaining critical sections that cannot be preempted are short and perform deterministically. RT scheduling latencies have been improved by three orders of magnitude and can now be measured roughly in tens of microseconds. Almost all interrupt handlers are converted to kernel threads that run in process context. Latency is lower and more deterministic because handlers become user-configurable, schedulable entities that can be preempted and prioritized just like any other process. High-resolution time and timers provide increased resolution and accuracy. RT Java uses these features for high-resolution sleep and timed waits. Linux high-resolution timers are implemented with a high-precision, 64-bit data type. Unlike traditional Linux, where time and timers depend on the low-resolution system tick -- which limits the granularity of timer events -- RT Linux uses independently programmable high-resolution timer events that can be made to expire within microseconds of each other. RT-Linux prevents priority Inversion by using priority inheritance. Following slides contains more information about this topic.
• #19: Java RTS real-time components and non-real-time components can coexist and share data on a single system. An RTS enabled JVM is fully compatible with Java SE giving developers unforeseen flexibility.
• #21: Due to these constrain Java is not suitable for real-time applications. One alternative may be codding real-time part of the application with another language, like C, and code the remaining non-real-time parts using Java. This technique is widely used in the past.
• #22: A Java-conformant JVM must delay loading a class until it's first referenced by a program. Loading a class can take a variable amount of time depending on the speed of the medium (disk or other) the class is loaded from, the class's size, and the overhead incurred by the class loaders themselves. If tens or hundreds of classes need to be loaded, the loading time itself can cause a significant and possibly unexpected delay. Careful application design can be used to load all classes at application start-up, but this must be done manually because the Java language specification doesn't let the JVM perform this step early.
• #23: When java was first introduced it was a interpreted language. But today nearly all JVMs use mixed mode execution. That means methods are initially interpreted but when the runtime has enough information about the code it is compiled to native code. HotSpot may even compile a method more than once to improve its performance when more information about the code is available. Althoug this behavior maximizes throughput in a server application (where java is dominated today) introduce significant nondeterminism about performance of a method. There are java compilers that produce native code directly for example gcj but moderm non-rt JVMs prefer JIT compilation because this makes lots of optimizations possible when compared to static compilation. For example polymorphic call to monomorphic call optimization, escape analyses.
• #24: Errors introduced by the need to manage memory explicitly in languages such as C and C++ are some of the most difficult problems to diagnose. Proving the absence of such errors when an application is deployed is also a fundamental challenge. One of the Java programming model's major strengths is that the JVM, not the application, performs memory management, which eliminates this burden for the application programmer. On the other hand, traditional garbage collectors can introduce long delays at times that are virtually impossible for the application programmer to predict. Delays of several hundred milliseconds are not unusual
• #25: Concurrent means works concurrently with mutator threads Parallel means more than one thread is used for collection Moving :explain fragmentation Generational:explain yound and old generation. Why this is necessary. Young generation collection and old generation collection. Fast/slow etc.
• #26: Different collectors can be used for different generation Parallel-scavenging:like serial collector but uses multiple threads.Only works on young generation Parallel compacting:like parallel-scavenging but works on yound and old generation CMS works on only old generation. Parallel GC is used for young generation. It is not a compacting collector. To combat fragmentation assumes feature object size demands based on past allocations. Some malloc implementations also combats fragmentation using simillar techniques.
• #28: Standard Java provides no guarantees for thread scheduling or thread priorities. An application that must respond to events in a well-defined time has no way to ensure that another low-priority thread won't get scheduled in front of a high-priority thread. Asynchronously Interrupted exceptions provides a safe way to stop a thread from another thread. But unlike Thread.stop this method stops the target thread in a controlled way. Priority Inversion may occur in non-rt operating systems and standard java. RTSJ requires an operating system that implements Priority Inheritance but allows others like Priority Ceiling emulation.
• #29: Java Real-Time Specification is shortly written as RTSJ. In 1998 a group of experts in real-time computer control was formed and coordinated by the National Institute of Standards and Technology (NIST) to draft requirements for real-time Java. The members were representatives from 50 different companies, government institutes and researchers from academic institutions. The main aim of the group was to develop a cross-disciplinary specification for real-time functionality that is expected to be needed by real-time applications written in Java programming language and being executed on various platforms. The group workshops at NIST produced nine core requirements for a Real-Time Java specification, together with number of derived sub-core requirements. This is the basis of the Real-Time Specification for Java (RTSJ). The weaknesses of the Java language specification for writing real-time applications are addressed in the core requirements.
• #31: If your application can tolerate some degree of non-determinism, then use a non-real-time virtual machine and tune it to obtain the response time you need, down to perhaps a 20 millisecond latency or so, but be aware that there will be occasional response time outliers. If you need response times below 20 milliseconds, even down to 30–70 microseconds, use Java RTS and its RTGC. Achieving any degree of predictability requires trading off application throughput in various ways. Virtual machine selection and configuration therefore occurs along a predictability spectrum. As virtual machines evolve, we can expect to be able to specify a desired level of predictability and have the system configure itself automatically to achieve it. Until then, a degree of manual configuration will be necessary. Using Java RTS and RTGC minimizes the effort necessary to achieve sub-millisecond levels of predictability.
• #35: It is a non-generational garbage collector. It operates on all heap at each execution. It has 3 modes of execution. It’s has 3 modes of execution. In normal mode priority is lowest RT priority by default (can change this using –XX:RTGCNormalPriority). It works concurrently with other threads. If free memory goes below a certain threshold (can be configured with –XX:NormalMinFreeBytes), it switches to boosted mode. In this mode its priority increased to boosted priority (Can be changed with –XX:RTGCBoostedPriority). Although in this mode GC can preempt RT Threads with priority lower than GC boosted priority, it can still work concurrently with them. If free memory continues to drop to a critical threshold, GC switches to critical mode. This mode is called deterministic mode. In this mode its priority is still boosted mode priority but all non-rt threads and threads with priority lower than GC boosted priority are suspended. If a thread wich has priority higher than boosted GC priority makes memory allocation it can be served from a special region called critical reserved memory (size can be adjusted using –XX:RTGCCriticalBoundary). Number of GC threads in normal and boosted/critical mode can be adjusted using commandline switches.
• #36: To use AOT or ITC compilation you have to provide list of classes and methods to be compiled ahead of time or at jvm startup. RTSJ implementations may provide tools to help you to generate this list.
• #37: A High Resolution Timer is required for an RT System. RT-Posix requires one. If you look at getTime(AbsoluteTime) method you can see that this method enables you to get time without creating a new AbsoluteTime object. Reason for this will be clear when we learn other features of RTSJ.
• #38: AbsoluteTime absolute(Clock clock) AbsoluteTime absolute(Clock clock, AbsoluteTime dest) int compareTo(HighResolutionTime time) int compareTo(java.lang.Object object) boolean equals(java.lang.Object object) boolean equals(HighResolutionTime object) long getMilliseconds() int getNanoseconds() int hashCode() RelativeTime relative(Clock clock) RelativeTime relative(Clock clock, RelativeTime time) void set(HighResolutionTime time) void set(long millis) void set(long millis, int nanos) static void waitForObject(java.lang.Object target, HighResolutionTime time) throws InterruptedException
• #40: Fixed priority preemptive scheduler does not change priority of the tasks automatically for this reason it is called fixed priority. Preemptive scheduler may preemt a task if a higher priority task is runnable.
• #41: Note that RT-Posix required 32 unique priorities. These requirements are also required by RT-Posix
• #42: Explain priority inversion by drawing on the board. The trouble experienced by the Mars lander " Mars Pathfinder " [1] [2] is a classic example of problems caused by priority inversion in realtime systems. Software continously reseted by watchdog timer due to priority inversion in actual mission. Error occured once in test phase but can not be repeated and clasified as a hardware glitch. Under the policy of priority inheritance, whenever a high priority task has to wait for some resource shared with an executing low priority task, the low priority task is assigned the priority of the highest waiting priority task for the duration of its own use of the shared resource, thus keeping medium priority tasks from pre-empting the (originally) low priority task With priority ceilings, the shared mutex process (that runs the operating system code) has a characteristic (high) priority of its own, which is assigned to the task locking the mutex. This works well, provided the other high priority task(s) that try to access the mutex does not have a priority higher than the ceiling priority.
• #43: T1:Low priority task is executing T2:Low priority task acquire shared lock T3:High priority task preempts low priority task T4:High priority task requests the same shared resource with low priority one and suspended. Low priority Task executes T5:Medium priority task preempts low priority task T6:Medium priority task continues to execute T7:Medium priority task executes and completes T8:Low priority task executes and completes T9:High priority task acuires the lock and completes
• #44: Real time threads are very similar to JLTs with some extra behaviors and constraints. They can do everything a JLT can do. You can specify several parameters parameters while creating them or at a later time. Most important parameters are as follows: SchedulingParameters: This parameter is handled by underlying scheduler and tell scheduler how this Thread should be scheduled. ReleaseParameters:As you will see at later slides, a thread can be periodic, aperiodic or sporadic. MemoryArea: RTSJ defines other memory regios than heap. A real-time thread can perform allocation in these new memory regions. The memory region passed with this parameter will be default allocation context of this thread. All parameters has a default value and null can be passed.
• #45: RTSJ defines a plugable scheduler architecture. Different schedulers may be plugged and each different scheduler may need different information as its scheduling parameters. RTSJ’s default priority scheduler defines scheduling parameter as priority so a PriorityParameters class is included to use with priority scheduler. Sometimes you may need to define an order between the task with the same priorities. Using ImportanceParameters you can define relative importance of tasks with the same priority. Base PriorityScheduler does NOT considers ImportanceParameters. Number of available priorities depends on target operating system. You can expect minimum of 28 different priorities. On solaris there are 60, on linux there are 49.
• #46: Periodic tasks occursat a known rate. It is easy to analyse feasibilty of meeting deadlines if you have all peridic task with known costs Aperiodic tasks does not have any occurance rate. They may occur at any time. Sporadic tasks are like aperiodic tasks but they have a known minimum inter occurance time. Aperidic taks can be scheduled using a sporadic server task. Which executes at a scheduled rate and executes a periodic taks.
• #47: Release parameters determine whether a thread is periodic, aperiodic or sporadic. You can specify time with nanoseconds resolution. Actual resolution will be platform dependent and you should test on your target environment. For a periodic task there is a minimum period that can be specified. This value is plaform dependent and should be tested. Your target environment may allow you to change this value to a lower one. All aperiodic events are put to a queue. Size of this queue can be specified using set Initial ArrivalTimeQueue Length method. You can also specify what to do if this queue fulls. Valid optionas are throwing an exception, disgarding event, increasing queue size accordingly or replacing the last one on the queue. Sporadic tasks are like aperiodic tasks. Instead this time you specify what to do task which does not obey their minimum inter arrival time using setMitViolationBehavior. Your valid options are the same.
• #49: A loop that uses sleep to execute once per period will cycle too slowly since every iteration of the loop uses sleep time plus the time to execute the code in the loop. Even measuring the time of the code in loop may not help. It is nearly impossible to get correct.
• #50: If a thread misses its deadline due to system overload executing an extra event handler may overload system more and cause extra deadline misses. You should design your miss handlers as light-weight as possible. Note that execution of deadline miss handler is not synchronous. Miss handler may block the thread, or may work parallel to thread depending on the priority of the handler and number of CPUs.A thread that misses a deadline continues to execute until it calls waitForNextPeriod. If thread does not have a deadline miss handler, waitForNextPeriod returns immediately with false value to signal a deadline miss. An RT Thread wihch misses a deadline is automatically descheduled, its deadline miss handler may reschedule it, terminate it using interruption.
• #51: Since NHRT are not allowed to access heap, they can preempt garbage collector at any time without any additional latency. Although normal RT Threads can also preempt GC they should always wait GC to be at a safe preemption point. They can not be created from normal threads because normal threads can not execute in scoped or immortal memory. They can only use ImmortalMemory or ScopedMemory. If a NHRT tries to access heap, MemoryAccessError is thrown. Enforcing this rule at runtime introduce additional runtime overhead. But some of it can be avoided by JVM by analysing application at runtime.
• #52: Sometimes hard real-time and non-realtime parts of the application need to change data. The important thing is not blocking a real-time thread due to a non real-time thread. To faciliate this wait free queues are provided. The write method appends a new element onto the queue. It is not synchronized, and does not block when the queue is full (it returns false instead). Multiple writer threads or schedulable objects are permitted, but if two or more threads intend to write to the same WaitFreeWriteQueue they will need to arrange explicit synchronization. If write is suscessfull write method returns true. Otherwise it returns null. Force method is similar to write method. The difference is if queue is full, it overrides the lates element in the queue. Return value of force method show whether an element is overriden. The read method removes the oldest element from the queue. It is synchronized, and will block when the queue is empty. It may be called by more than one reader, in which case the different callers will read different elements from the queue.
• #53: The write method appends a new element onto the queue. It is synchronized, and blocks when the queue is full. It may be called by more than one writer, in which case, the different callers will write to different elements of the queue. The read method removes the oldest element from the queue. It is not synchronized and does not block; it will return null when the queue is empty.Multiple reader threads or schedulable objects are permitted, but if two or more intend to read from the same WaitFreeWriteQueue they will need to arrange explicit synchronization.
• #54: Each schedulable has an allocation context. It can be one of the 4 defined main memory regions. ImmortalMemory is a memory resource that is unexceptionally available to all schedulable objects and Java threads for use and allocation. An immortal object may not contain references to any form of scoped memory. Object in immortalMemory are never garbage collected. All static data goes into this region. Objects allocated in scoped memory are freed when (and only when) no schedulable object has access to the objects in the scoped memory. When a ScopedMemory area is instantiated, the object itself is allocated from the current memory allocation context, but the memory space that object represents (it's backing store) is allocated from memory that is not otherwise directly visible to Java code; e.g., it might be allocated with the C malloc function. This backing store behaves effectively as if it were allocated when the associated scoped memory object is constructed and freed at that scoped memory object's finalization. The enter() method of ScopedMemory is one mechanism used to make a memory area the current allocation context. The other mechanism for activating a memory area is making it the initial memory area for a real-time thread or async event handler. An instance of ImmortalPhysicalMemory allows objects to be allocated from a range of physical memory with particular attributes, determined by their memory type. This memory area has the same restrictive set of assignment rules as ImmortalMemory memory areas, and may be used in any execution context where ImmortalMemory is appropriate. LTMemory represents a memory area guaranteed by the system to have linear time allocation when memory consumption from the memory area is less than the memory area's initial size. Execution time for allocation is allowed to vary when memory consumption is between the initial size and the maximum size for the area. Furthermore, the underlying system is not required to guarantee that memory between initial and maximum will always be available. Objects in LTMemory can be safely accessed from NoHeapRealTimeThread. VTMemory is similar to LTMemory except that the execution time of an allocation from a VTMemory area need not complete in linear time.
• #55: MemoryArea is the abstract base class of all classes dealing with the representations of allocatable memory areas, including the immortal memory area, physical memory and scoped memory areas.
• #56: An instance of RawMemoryAccess models a range of physical memory as a fixed sequence of bytes. A full complement of accessor methods allow the contents of the physical area to be accessed through offsets from the base, interpreted as byte, short, int, or long data values or as arrays of these types. The RawMemoryAccess class allows a real-time program to implement device drivers, memory-mapped I/O, flash memory, battery-backed RAM, and similar low-level software. List of methods in the diagram are not complete.
• #58: How binding is performed is not specified in RTSJ. A bind method is defined but possible values of its string argument are not defined and platform specific. You may create your own event objects by extending from AsyncEvent. They are not only for external events. AEHs are designed to execute a few lines of code, then exit. They are not expected to block, sleep, or change its scheduling parameters.
• #61: 1. You don't need to guess the capacity of a TextBuilder, FastTable or a FastMap, their size expand gently without ever incurring expensive resize/copy or rehash operations (unlike StringBuilder, ArrayList or HashMap). 3. Javolution classes are fast, very fast (e.g. Text insertion/deletion in O[Log(n)] instead of O[n] for standard StringBuffer/StringBuilder). 4. All Javolution classes are hard real-time compliant and have highly deterministic behavior (in the microsecond range). Furthermore (unlike the standard library), Javolution is RTSJ safe (no memory clash or memory leak when used with Java Real-Time extension). 5. Javolution makes it easy for concurrent algorithms to take advantage of multi-processors systems. 6. Javolution's real-time collection classes (map, list, table and set) can be used in place of most standard collection classes and provide numerous additional capabilities. 7. Any Java class can be serialized/deserialized in XML format in any form you may want, also no need to implement Serializable or for the platform to support serialization 8. Javolution provides Struct and Union classes for direct interoperability with C/C++ applications. 9. Javolution runs on any platform from the simplest J2ME CLDC 1.0 with no garbage collector to the latest J2EE 5.0 with parameterized types. 10. Javolution is a pure Java Solution (no native code), small (less than 300 KBytes jar file) and free; permission to use, copy, modify, and distribute this software is freely granted, provided that copyright notices are preserved (BSD License).
• #63: The reason for scheduling against time instead of allocation rate is that allocation is often uneven during an application's execution. Metronome divides time into a series of discrete quanta, approximately 500 microseconds but no more than 1 millisecond in length, that are devoted to either GC work or application work. Arraylets break up large arrays into smaller pieces to make large array allocations easier to satisfy without defragmenting the heap. The arraylet object's first level, known as the spine, contains a list of pointers to the array's smaller pieces, known as leaves. Each leaf is the same size, which simplifies the calculation to find any particular element of the array and also makes it easier for the collector to find a suitable free space to allocate each leaf. Breaking arrays up into smaller noncontiguous pieces lets arrays be allocated within the many smaller free areas that typically occur on a heap, without needing to compact.
• #65: All the classes in a jar file can be compiled or only hot methods can be compiled using information obtained from a previous run of the application. Although AOT code enables more-deterministic performance, it also has some disadvantages. The JXEs used to store AOT code are generally much larger than the JAR files that hold the class files because native code is generally less dense than the bytecodes stored in class files.A second disadvantage is that AOT-compiled code, though faster than interpreted code, can be substantially slower than JIT-compiled code. To avoid nondeterministic performance effects, neither the JIT compiler nor the AOT compiler provided in WebSphere Real Time applies the aggressively speculative optimizations generally applied by modern JIT compilers.
• #74: Java byte code will be converted to target platform prior to deployment and certification. Most probably garbage collector will not be available because it is very complex to certify and not needed by this kind of applications where resources are statically allocated. A set of tools for analysing source code for verification will be provided
• #75: Aonix safety critical JVM is based on JSR 302. Selected to be evaluated in DIANA project that aims to evaluate new tools to modernize tool chains used in avionics. Aonix safety critical JVM (Raven) will be also evaluated in Taranis, a technology demonstration program for enhanced aerial vehicles for the UK Ministry of De fence