Chapter 4 PDF - Process Management

This chapter examines some more advanced concepts related to process management, which are found in a number of contemporary operating systems. We show that the concept of process is more complex and subtle than presented so far and in fact embodies two separate and potentially independent concep...

This chapter examines some more advanced concepts related to process management, which are found in a number of contemporary operating systems. We show that the concept of process is more complex and subtle than presented so far and in fact embodies two separate and potentially independent concepts: one relating to resource ownership, and another relating to execution. This distinction has led to the development, in many operating systems, of a construct known as the thread. 4.1 PROCESSES AND THREADS The discussion so far has presented the concept of a process as embodying two characteristics: 1. Resource ownership: A process includes a virtual address space to hold the process image; recall from Chapter 3 that the process image is the collection of program, data, stack, and attributes defined in the process control block. From time to time, a process may be allocated control or ownership of resources, such as main memory, I/O channels, I/O devices, and files. The OS performs a protection function to prevent unwanted interference between processes with respect to resources. 2. Scheduling/execution: The execution of a process follows an execution path (trace) through one or more programs (e.g., Figure 1.5). This execution may be interleaved with that of other processes. Thus, a process has an execution state (Running, Ready, etc.) and a dispatching priority, and is the entity that is scheduled and dispatched by the OS. Some thought should convince the reader that these two characteristics are independent and could be treated independently by the OS. This is done in a number of operating systems, particularly recently developed systems. To distinguish the two characteristics, the unit of dispatching is usually referred to as a thread or lightweight process, while the unit of resource ownership is usually referred to as a process or task. Multithreading Multithreading refers to the ability of an OS to support multiple, concurrent paths of execution within a single process. The traditional approach of a single thread of execution per process, in which the concept of a thread is not recognized, is referred to as a single-threaded approach. The two arrangements shown in the left half of Figure 4.1 are single-threaded approaches. MS-DOS is an example of an OS that supports a single-user process and a single thread. Other operating systems, such as some variants of UNIX, support multiple user processes, but only support one thread per process. The right half of Figure 4.1 depicts multithreaded approaches. A Java runtime environment is an example of a system of one process with multiple threads. Of interest in this section is the use of multiple processes, each of which supports multiple threads. This approach is taken in Windows, Solaris, and many 1Alas, even this degree of consistency is not maintained. In IBM’s mainframe operating systems, the concepts of address space and task, respectively, correspond roughly to the concepts of process and thread that we describe in this section. Also, in the literature, the term lightweight process is used as either (1) equivalent to the term thread, (2) a particular type of thread known as a kernel-level thread, or (3) in the case of Solaris, an entity that maps user-level threads to kernel-level threads. Figure 4.1 Threads and Processes One process One thread One process Multiple threads Multiple processes One thread per process = Instruction trace Multiple processes Multiple threads per process 4.1 / PROCESSES AND THREADS 179 modern versions of UNIX, among others. In this section, we give a general description of multithreading; the details of the Windows, Solaris, and Linux approaches will be discussed later in this chapter. In a multithreaded environment, a process is defined as the unit of resource allocation and a unit of protection. The following are associated with processes: A virtual address space that holds the process image Protected access to processors, other processes (for interprocess communication), files, and I/O resources (devices and channels) Within a process, there may be one or more threads, each with the following: A thread execution state (Running, Ready, etc.) A saved thread context when not running; one way to view a thread is as an independent program counter operating within a process An execution stack Some per-thread static storage for local variables Access to the memory and resources of its process, shared with all other threads in that process Figure 4.2 illustrates the distinction between threads and processes from the point of view of process management. In a single-threaded process model (i.e., there is no distinct concept of thread), the representation of a process includes its process control block and user address space, as well as user and kernel stacks to manage the call/return behavior of the execution of the process. While the process is running, it controls the processor registers. The contents of these registers are saved when the process is not running. In a multithreaded environment, there is still a single process Figure 4.2 Single-Threaded and Multithreaded Process Models Single-threaded process model Process control block User address space User stack Kernel stack Multithreaded process model Process control block User address space User stack Kernel stack User stack Kernel stack User stack Kernel stack Thread control block Thread Thread Thread Thread control block Thread control block 180 Chapter 4 / Threads control block and user address space associated with the process, but now there are separate stacks for each thread, as well as a separate control block for each thread containing register values, priority, and other thread-related state information. Thus, all of the threads of a process share the state and resources of that process. They reside in the same address space and have access to the same data. When one thread alters an item of data in memory, other threads see the results if and when they access that item. If one thread opens a file with read privileges, other threads in the same process can also read from that file. The key benefits of threads derive from the performance implications: 1. It takes far less time to create a new thread in an existing process, than to create a brand-new process. Studies done by the Mach developers show that thread creation is ten times faster than process creation in UNIX [TEVA87]. 2. It takes less time to terminate a thread than a process. 3. It takes less time to switch between two threads within the same process than to switch between processes. 4. Threads enhance efficiency in communication between different executing programs. In most operating systems, communication between independent processes requires the intervention of the kernel to provide protection and the mechanisms needed for communication. However, because threads within the same process share memory and files, they can communicate with each other without invoking the kernel. Thus, if there is an application or function that should be implemented as a set of related units of execution, it is far more efficient to do so as a collection of threads, rather than a collection of separate processes. An example of an application that could make use of threads is a file server. As each new file request comes in, a new thread can be spawned for the file management program. Because a server will handle many requests, many threads will be created and destroyed in a short period. If the server runs on a multiprocessor computer, then multiple threads within the same process can be executing simultaneously on different processors. Further, because processes or threads in a file server must share file data and therefore coordinate their actions, it is faster to use threads and shared memory than processes and message passing for this coordination. The thread construct is also useful on a single processor to simplify the structure of a program that is logically doing several different functions. [LETW88] gives four examples of the uses of threads in a single-user multiprocessing system: 1. Foreground and background work: For example, in a spreadsheet program, one thread could display menus and read user input, while another thread executes user commands and updates the spreadsheet. This arrangement often increases the perceived speed of the application by allowing the program to prompt for the next command before the previous command is complete. 2. Asynchronous processing: Asynchronous elements in the program can be implemented as threads. For example, as a protection against power failure, one can design a word processor to write its random access memory (RAM) 4.1 / PROCESSES AND THREADS 181 buffer to disk once every minute. A thread can be created whose sole job is periodic backup and that schedules itself directly with the OS; there is no need for fancy code in the main program to provide for time checks or to coordinate input and output. 3. Speed of execution: A multithreaded process can compute one batch of data while reading the next batch from a device. On a multiprocessor system, multiple threads from the same process may be able to execute simultaneously. Thus, even though one thread may be blocked for an I/O operation to read in a batch of data, another thread may be executing. 4. Modular program structure: Programs that involve a variety of activities or a variety of sources and destinations of input and output may be easier to design and implement using threads. In an OS that supports threads, scheduling and dispatching is done on a thread basis; hence, most of the state information dealing with execution is maintained in thread-level data structures. There are, however, several actions that affect all of the threads in a process, and that the OS must manage at the process level. For example, suspension involves swapping the address space of one process out of main memory to make room for the address space of another process. Because all threads in a process share the same address space, all threads are suspended at the same time. Similarly, termination of a process terminates all threads within that process. Thread Functionality Like processes, threads have execution states and may synchronize with one another. We look at these two aspects of thread functionality in turn. Thread States As with processes, the key states for a thread are Running, Ready, and Blocked. Generally, it does not make sense to associate suspend states with threads because such states are process-level concepts. In particular, if a process is swapped out, all of its threads are necessarily swapped out because they all share the address space of the process. There are four basic thread operations associated with a change in thread state [ANDE04]: 1. Spawn: Typically, when a new process is spawned, a thread for that process is also spawned. Subsequently, a thread within a process may spawn another thread within the same process, providing an instruction pointer and arguments for the new thread. The new thread is provided with its own register context and stack space and placed on the Ready queue. 2. Block: When a thread needs to wait for an event, it will block (saving its user registers, program counter, and stack pointers). The processor may then turn to the execution of another ready thread in the same or a different process. 3. Unblock: When the event for which a thread is blocked occurs, the thread is moved to the Ready queue. 4. Finish: When a thread completes, its register context and stacks are deallocated. 182 Chapter 4 / Threads A significant issue is whether the blocking of a thread results in the blocking of the entire process. In other words, if one thread in a process is blocked, does this prevent the running of any other thread in the same process, even if that other thread is in a ready state? Clearly, some of the flexibility and power of threads is lost if the one blocked thread blocks an entire process. We will return to this issue subsequently in our discussion of user-level versus kernel-level threads, but for now, let us consider the performance benefits of threads that do not block an entire process. Figure 4.3 (based on one in [KLEI96]) shows a program that performs two remote procedure calls (RPCs)2 to two different hosts to obtain a combined result. In a single-threaded program, the results are obtained in sequence, so the program has to wait for a response from each server in turn. Rewriting the program to use a separate thread for each RPC results in a substantial speedup. Note if this program operates on a uniprocessor, the requests must be generated sequentially and the results processed in sequence; however, the program waits concurrently for the two replies. 2An RPC is a technique by which two programs, which may execute on different machines, interact using procedure call/return syntax and semantics. Both the called and calling programs behave as if the partner program were running on the same machine. RPCs are often used for client/server applications and will be discussed in Chapter 16. Figure 4.3 Remote Procedure Call (RPC) Using Threads (a) RPC using single thread (b) RPC using one thread per server (on a uniprocessor) Time Process 1 Blocked, waiting for response to RPC Blocked, waiting for processor, which is in use by Thread B Running Thread A (Process 1) Thread B (Process 1) Server Server Server Server RPC request RPC request RPC request RPC request 4.2 / TYPES OF THREADS 183 On a uniprocessor, multiprogramming enables the interleaving of multiple threads within multiple processes. In the example of Figure 4.4, three threads in two processes are interleaved on the processor. Execution passes from one thread to another either when the currently running thread is blocked or when its time slice is exhausted.3 Thread Synchronization All of the threads of a process share the same address space and other resources, such as open files. Any alteration of a resource by one thread affects the environment of the other threads in the same process. It is therefore necessary to synchronize the activities of the various threads so that they do not interfere with each other or corrupt data structures. For example, if two threads each try to add an element to a doubly linked list at the same time, one element may be lost or the list may end up malformed. The issues raised and the techniques used in the synchronization of threads are, in general, the same as for the synchronization of processes. These issues and techniques will be the subject of Chapters 5 and 6. 4.2 TYPES OF THREADS User-Level and Kernel-Level Threads There are two broad categories of thread implementation: user-level threads (ULTs) and kernel-level threads (KLTs).4 The latter are also referred to in the literature as kernel-supported threads or lightweight processes. User-Level Threads In a pure ULT facility, all of the work of thread management is done by the application and the kernel is not aware of the existence of threads. 3In this example, thread C begins to run after thread A exhausts its time quantum, even though thread B is also ready to run. The choice between B and C is a scheduling decision, a topic covered in Part Four. 4The acronyms ULT and KLT are not widely used, but are introduced for conciseness. Figure 4.4 Multithreading Example on a Uniprocessor Time Blocked I/O request Thread A (Process 1) Thread B (Process 1) Thread C (Process 2) Ready Running Request complete Time quantum expires Time quantum expires Process created 184 Chapter 4 / Threads Figure 4.5a illustrates the pure ULT approach. Any application can be programmed to be multithreaded by using a threads library, which is a package of routines for ULT management. The threads library contains code for creating and destroying threads, for passing messages and data between threads, for scheduling thread execution, and for saving and restoring thread contexts. By default, an application begins with a single thread and begins running in that thread. This application and its thread are allocated to a single process managed by the kernel. At any time that the application is running (the process is in the Running state), the application may spawn a new thread to run within the same process. Spawning is done by invoking the spawn utility in the threads library. Control is passed to that utility by a procedure call. The threads library creates a data structure for the new thread and then passes control to one of the threads within this process that is in the Ready state, using some scheduling algorithm. When control is passed to the library, the context of the current thread is saved, and when control is passed from the library to a thread, the context of that thread is restored. The context essentially consists of the contents of user registers, the program counter, and stack pointers. All of the activity described in the preceding paragraph takes place in user space and within a single process. The kernel is unaware of this activity. The kernel continues to schedule the process as a unit and assigns a single execution state (Ready, Running, Blocked, etc.) to that process. The following examples should clarify the relationship between thread scheduling and process scheduling. Suppose process B is executing in its thread 2; the states of the process and two ULTs that are part of the process are shown in Figure 4.6a. Each of the following is a possible occurrence: 1. The application executing in thread 2 makes a system call that blocks B. For example, an I/O call is made. This causes control to transfer to the kernel. The kernel invokes the I/O action, places process B in the Blocked state, and Figure 4.5 User-Level and Kernel-Level Threads PP User space Threads library Kernel space P P User space Kernel space P User space Threads library Kernel space (a) Pure user–level (b) Pure kernel–level (c) Combined User-level thread Kernel-level thread Process Figure 4.6 Examples of the Relationships between User-Level Thread States and Process States Ready Running Blocked Thread 1 Ready Running Blocked Thread 2 Ready Running Process B (a) Ready Running Blocked Thread 1 Ready Running Blocked Thread 2 Ready Running Process B (b) Ready Running Blocked Blocked Thread 1 Ready Running Blocked Thread 2 Ready Running Blocked Process B (c) Ready Running Blocked Thread 1 Ready Running Blocked Blocked Thread 2 Ready Running Blocked Process B (d) 185 186 Chapter 4 / Threads switches to another process. Meanwhile, according to the data structure maintained by the threads library, thread 2 of process B is still in the Running state. It is important to note that thread 2 is not actually running in the sense of being executed on a processor; but it is perceived as being in the Running state by the threads library. The corresponding state diagrams are shown in Figure 4.6b. 2. A clock interrupt passes control to the kernel, and the kernel determines that the currently running process (B) has exhausted its time slice. The kernel places process B in the Ready state and switches to another process. Meanwhile, according to the data structure maintained by the threads library, thread 2 of process B is still in the Running state. The corresponding state diagrams are shown in Figure 4.6c. 3. Thread 2 has reached a point where it needs some action performed by thread 1 of process B. Thread 2 enters a Blocked state and thread 1 transitions from Ready to Running. The process itself remains in the Running state. The corresponding state diagrams are shown in Figure 4.6d. Note that each of the three preceding items suggests an alternative event starting from diagram (a) of Figure 4.6. So each of the three other diagrams (b, c, d) shows a transition from the situation in (a). In cases 1 and 2 (Figures 4.6b and 4.6c), when the kernel switches control back to process B, execution resumes in thread 2. Also note that a process can be interrupted, either by exhausting its time slice or by being preempted by a higher-priority process, while it is executing code in the threads library. Thus, a process may be in the midst of a thread switch from one thread to another when interrupted. When that process is resumed, execution continues within the threads library, which completes the thread switch and transfers control to another thread within that process. There are a number of advantages to the use of ULTs instead of KLTs, including the following: 1. Thread switching does not require kernel-mode privileges because all of the thread management data structures are within the user address space of a single process. Therefore, the process does not switch to the kernel mode to do thread management. This saves the overhead of two mode switches (user to kernel; kernel back to user). 2. Scheduling can be application specific. One application may benefit most from a simple round-robin scheduling algorithm, while another might benefit from a priority-based scheduling algorithm. The scheduling algorithm can be tailored to the application without disturbing the underlying OS scheduler. 3. ULTs can run on any OS. No changes are required to the underlying kernel to support ULTs. The threads library is a set of application-level functions shared by all applications. There are two distinct disadvantages of ULTs compared to KLTs: 1. In a typical OS, many system calls are blocking. As a result, when a ULT executes a system call, not only is that thread blocked, but all of the threads within the process are blocked as well. 4.2 / TYPES OF THREADS 187 2. In a pure ULT strategy, a multithreaded application cannot take advantage of multiprocessing. A kernel assigns one process to only one processor at a time. Therefore, only a single thread within a process can execute at a time. In effect, we have application-level multiprogramming within a single process. While this multiprogramming can result in a significant speedup of the application, there are applications that would benefit from the ability to execute portions of code simultaneously. There are ways to work around these two problems. For example, both problems can be overcome by writing an application as multiple processes rather than multiple threads. But this approach eliminates the main advantage of threads: Each switch becomes a process switch rather than a thread switch, resulting in much greater overhead. Another way to overcome the problem of blocking threads is to use a technique referred to as jacketing. The purpose of jacketing is to convert a blocking system call into a nonblocking system call. For example, instead of directly calling a system I/O routine, a thread calls an application-level I/O jacket routine. Within this jacket routine is code that checks to determine if the I/O device is busy. If it is, the thread enters the Blocked state and passes control (through the threads library) to another thread. When this thread is later given control again, the jacket routine checks the I/O device again. Kernel-Level Threads In a pure KLT facility, all of the work of thread management is done by the kernel. There is no thread management code in the application level, simply an application programming interface (API) to the kernel thread facility. Windows is an example of this approach. Figure 4.5b depicts the pure KLT approach. The kernel maintains context information for the process as a whole and for individual threads within the process. Scheduling by the kernel is done on a thread basis. This approach overcomes the two principal drawbacks of the ULT approach. First, the kernel can simultaneously schedule multiple threads from the same process on multiple processors. Second, if one thread in a process is blocked, the kernel can schedule another thread of the same process. Another advantage of the KLT approach is that kernel routines themselves can be multithreaded. The principal disadvantage of the KLT approach compared to the ULT approach is that the transfer of control from one thread to another within the same process requires a mode switch to the kernel. To illustrate the differences, Table 4.1 shows the results of measurements taken on a uniprocessor VAX computer running a UNIX-like OS. The two benchmarks are as follows: Null Fork, the time to create, schedule, execute, and complete a process/thread that invokes Operation User-Level Threads Kernel-Level Threads Processes Null Fork 34 948 11,300 Signal Wait 37 441 1,840 Table 4.1 Thread and Process Operation Latencies (ms) 188 Chapter 4 / Threads the null procedure (i.e., the overhead of forking a process/thread); and Signal-Wait, the time for a process/thread to signal a waiting process/thread and then wait on a condition (i.e., the overhead of synchronizing two processes/threads together). We see there is an order of magnitude or more of difference between ULTs and KLTs, and similarly between KLTs and processes. Thus, on the face of it, while there is a significant speedup by using KLT multithreading compared to single-threaded processes, there is an additional significant speedup by using ULTs. However, whether or not the additional speedup is realized depends on the nature of the applications involved. If most of the thread switches in an application require kernel-mode access, then a ULT-based scheme may not perform much better than a KLT-based scheme. Combined Approaches Some operating systems provide a combined ULT/KLT facility (see Figure 4.5c). In a combined system, thread creation is done completely in user space, as is the bulk of the scheduling and synchronization of threads within an application. The multiple ULTs from a single application are mapped onto some (smaller or equal) number of KLTs. The programmer may adjust the number of KLTs for a particular application and processor to achieve the best overall results. In a combined approach, multiple threads within the same application can run in parallel on multiple processors, and a blocking system call need not block the entire process. If properly designed, this approach should combine the advantages of the pure ULT and KLT approaches while minimizing the disadvantages. Solaris is a good example of an OS using this combined approach. The current Solaris version limits the ULT/KLT relationship to be one-to-one. Other Arrangements As we have said, the concepts of resource allocation and dispatching unit have traditionally been embodied in the single concept of the process—that is, as a 1 : 1 relationship between threads and processes. Recently, there has been much interest in providing for multiple threads within a single process, which is a many-to-one relationship. However, as Table 4.2 shows, the other two combinations have also been investigated, namely, a many-to-many relationship and a one-to-many relationship. Threads: Processes Description Example Systems 1:1 Each thread of execution is a unique process with its own address space and resources. Traditional UNIX implementations M:1 A process defines an address space and dynamic resource ownership. Multiple threads may be created and executed within that process. Windows NT, Solaris, Linux, OS/2, OS/390, MACH 1:M A thread may migrate from one process environment to another. This allows a thread to be easily moved among distinct systems. Ra (Clouds), Emerald M:N It combines attributes of M:1 and 1:M cases. TRIX Table 4.2 Relationship between Threads and Processes 4.2 / TYPES OF THREADS 189 Many-to-Many Relationship The idea of having a many-to-many relationship between threads and processes has been explored in the experimental operating system TRIX [PAZZ92, WARD80]. In TRIX, there are the concepts of domain and thread. A domain is a static entity, consisting of an address space and “ports” through which messages may be sent and received. A thread is a single execution path, with an execution stack, processor state, and scheduling information. As with the multithreading approaches discussed so far, multiple threads may execute in a single domain, providing the efficiency gains discussed earlier. However, it is also possible for a single-user activity, or application, to be performed in multiple domains. In this case, a thread exists that can move from one domain to another. The use of a single thread in multiple domains seems primarily motivated by a desire to provide structuring tools for the programmer. For example, consider a program that makes use of an I/O subprogram. In a multiprogramming environment that allows user-spawned processes, the main program could generate a new process to handle I/O, then continue to execute. However, if the future progress of the main program depends on the outcome of the I/O operation, then the main program will have to wait for the other I/O program to finish. There are several ways to implement this application: 1. The entire program can be implemented as a single process. This is a reasonable and straightforward solution. There are drawbacks related to memory management. The process as a whole may require considerable main memory to execute efficiently, whereas the I/O subprogram requires a relatively small address space to buffer I/O and to handle the relatively small amount of program code. Because the I/O program executes in the address space of the larger program, either the entire process must remain in main memory during the I/O operation, or the I/O operation is subject to swapping. This memory management effect would also exist if the main program and the I/O subprogram were implemented as two threads in the same address space. 2. The main program and I/O subprogram can be implemented as two separate processes. This incurs the overhead of creating the subordinate process. If the I/O activity is frequent, one must either leave the subordinate process alive, which consumes management resources, or frequently create and destroy the subprogram, which is inefficient. 3. Treat the main program and the I/O subprogram as a single activity that is to be implemented as a single thread. However, one address space (domain) could be created for the main program and one for the I/O subprogram. Thus, the thread can be moved between the two address spaces as execution proceeds. The OS can manage the two address spaces independently, and no process creation overhead is incurred. Furthermore, the address space used by the I/O subprogram could also be shared by other simple I/O programs. The experiences of the TRIX developers indicate that the third option has merit, and may be the most effective solution for some applications. One-to-Many Relationship In the field of distributed operating systems (designed to control distributed computer systems), there has been interest in the 190 Chapter 4 / Threads concept of a thread as primarily an entity that can move among address spaces.5 A notable example of this research is the Clouds operating system, and especially its kernel, known as Ra [DASG92]. Another example is the Emerald system [STEE95]. A thread in Clouds is a unit of activity from the user’s perspective. A process is a virtual address space with an associated process control block. Upon creation, a thread starts executing in a process by invoking an entry point to a program in that process. Threads may move from one address space to another, and actually span computer boundaries (i.e., move from one computer to another). As a thread moves, it must carry with it certain information, such as the controlling terminal, global parameters, and scheduling guidance (e.g., priority). The Clouds approach provides an effective way of insulating both users and programmers from the details of the distributed environment. A user’s activity may be represented as a single thread, and the movement of that thread among computers may be dictated by the OS for a variety of system-related reasons, such as the need to access a remote resource, and load balancing. 4.3 MULTICORE AND MULTITHREADING The use of a multicore system to support a single application with multiple threads (such as might occur on a workstation, a video game console, or a personal computer running a processor-intense application) raises issues of performance and application design. In this section, we first look at some of the performance implications of a multithreaded application on a multicore system, then describe a specific example of an application designed to exploit multicore capabilities. Performance of Software on Multicore The potential performance benefits of a multicore organization depend on the ability to effectively exploit the parallel resources available to the application. Let us focus first on a single application running on a multicore system. Amdahl’s law (see Appendix E) states that: Speedup = time to execute program on a single processor time to execute program on N parallel processors = 1 (1 - f) + f N The law assumes a program in which a fraction (1 - f) of the execution time involves code that is inherently serial, and a fraction f that involves code that is infinitely parallelizable with no scheduling overhead. This law appears to make the prospect of a multicore organization attractive. But as Figure 4.7a shows, even a small amount of serial code has a noticeable impact. If only 10% of the code is inherently serial (f = 0.9), running the program on a multicore system with eight processors yields a performance gain of a factor of only 4.7. In 5The movement of processes or threads among address spaces, or thread migration, on different machines has become a hot topic in recent years. Chapter 18 will explore this topic. 4.3 / MULTICORE AND MULTITHREADING 191 addition, software typically incurs overhead as a result of communication and distribution of work to multiple processors and cache coherence overhead. This results in a curve where performance peaks and then begins to degrade because of the increased burden of the overhead of using multiple processors. Figure 4.7b, from [MCDO07], is a representative example. However, software engineers have been addressing this problem, and there are numerous applications in which it is possible to effectively exploit a multicore Figure 4.7 Performance Effect of Multiple Cores Relative speedup 0 2 4 6 8 12 Number of processors (a) Speedup with 0%, 2%, 5%, and 10% sequential portions 3456 0% 2% 5% 10% 78 Relative speedup 10% 5% 15% 20% 0 0.5 1.0 1.5 2.0 2.5 12 Number of processors (b) Speedup with overheads 345678 192 Chapter 4 / Threads system. [MCDO07] reports on a set of database applications, in which great attention was paid to reducing the serial fraction within hardware architectures, operating systems, middleware, and the database application software. Figure 4.8 shows the result. As this example shows, database management systems and database applications are one area in which multicore systems can be used effectively. Many kinds of servers can also effectively use the parallel multicore organization, because servers typically handle numerous relatively independent transactions in parallel. In addition to general-purpose server software, a number of classes of applications benefit directly from the ability to scale throughput with the number of cores. [MCDO06] lists the following examples: Multithreaded native applications: Multithreaded applications are characterized by having a small number of highly threaded processes. Examples of threaded applications include Lotus Domino or Siebel CRM (Customer Relationship Manager).  Multiprocess applications: Multiprocess applications are characterized by the presence of many single-threaded processes. Examples of multiprocess applications include the Oracle database, SAP, and PeopleSoft. Java applications: Java applications embrace threading in a fundamental way. Not only does the Java language greatly facilitate multithreaded applications, but the Java Virtual Machine is a multithreaded process that provides scheduling and memory management for Java applications. Java applications that can benefit directly from multicore resources include application servers such as Oracle’s Java Application Server, BEA’s Weblogic, IBM’s Websphere, and the open-source Tomcat application server. All applications that use a Java 2 Figure 4.8 Scaling of Database Workloads on Multiprocessor Hardware 0 0 16 32 48 64 16 32 Number of CPUs Scaling 48 64 perfect scaling Oracle DSS 4-way join TMC data mining DB2 DSS scan & aggs Oracle ad hoc insurance OLTP 4.3 / MULTICORE AND MULTITHREADING 193 Platform, Enterprise Edition (J2EE platform) application server can immediately benefit from multicore technology. Multi-instance applications: Even if an individual application does not scale to take advantage of a large number of threads, it is still possible to gain from multicore architecture by running multiple instances of the application in parallel. If multiple application instances require some degree of isolation, virtualization technology (for the hardware of the operating system) can be used to provide each of them with its own separate and secure environment. Application Example: Valve Game Software Valve is an entertainment and technology company that has developed a number of popular games, as well as the Source engine, one of the most widely played game engines available. Source is an animation engine used by Valve for its games and licensed for other game developers. In recent years, Valve has reprogrammed the Source engine software to use multithreading to exploit the power of multicore processor chips from Intel and AMD [REIM06]. The revised Source engine code provides more powerful support for Valve games such as Half Life 2. From Valve’s perspective, threading granularity options are defined as follows [HARR06]: Coarse threading: Individual modules, called systems, are assigned to individual processors. In the Source engine case, this would mean putting rendering on one processor, AI (artificial intelligence) on another, physics on another, and so on. This is straightforward. In essence, each major module is single-threaded and the principal coordination involves synchronizing all the threads with a timeline thread.  Fine-grained threading: Many similar or identical tasks are spread across multiple processors. For example, a loop that iterates over an array of data can be split up into a number of smaller parallel loops in individual threads that can be scheduled in parallel. Hybrid threading: This involves the selective use of fine-grained threading for some systems, and single-threaded for other systems. Valve found that through coarse threading, it could achieve up to twice the performance across two processors compared to executing on a single processor. But this performance gain could only be achieved with contrived cases. For real-world gameplay, the improvement was on the order of a factor of 1.2. Valve also found that effective use of fine-grained threading was difficult. The time-per-work unit can be variable, and managing the timeline of outcomes and consequences involved complex programming. Valve found that a hybrid threading approach was the most promising and would scale the best, as multicore systems with 8 or 16 processors became available. Valve identified systems that operate very effectively being permanently assigned to a single processor. An example is sound mixing, which has little user interaction, is not constrained by the frame configuration of windows, and works on its own set 194 Chapter 4 / Threads of data. Other modules, such as scene rendering, can be organized into a number of threads so that the module can execute on a single processor but achieve greater performance as it is spread out over more and more processors. Figure 4.9 illustrates the thread structure for the rendering module. In this hierarchical structure, higher-level threads spawn lower-level threads as needed. The rendering module relies on a critical part of the Source engine, the world list, which is a database representation of the visual elements in the game’s world. The first task is to determine what are the areas of the world that need to be rendered. The next task is to determine what objects are in the scene as viewed from multiple angles. Then comes the processor-intensive work. The rendering module has to work out the rendering of each object from multiple points of view, such as the player’s view, the view of monitors, and the point of view of reflections in water. Some of the key elements of the threading strategy for the rendering module are listed in [LEON07] and include the following: Construct scene rendering lists for multiple scenes in parallel (e.g., the world and its reflection in water). Overlap graphics simulation. Compute character bone transformations for all characters in all scenes in parallel. Allow multiple threads to draw in parallel. Figure 4.9 Hybrid Threading for Rendering Module Render Skybox Main view Scene list For each object Particles Sim and draw Bone setup Draw Character Etc. Monitor Etc. 4.4 / WINDOWS PROCESS AND THREAD MANAGEMENT 195 The designers found that simply locking key databases, such as the world list, for a thread was too inefficient. Over 95% of the time, a thread is trying to read from a data set, and only 5% of the time at most is spent in writing to a data set. Thus, a concurrency mechanism known as the single-writer-multiple-readers model works effectively. 4.4 WINDOWS PROCESS AND THREAD MANAGEMENT This section begins with an overview of the key objects and mechanisms that support application execution in Windows. The remainder of the section looks in more detail at how processes and threads are managed. An application consists of one or more processes. Each process provides the resources needed to execute a program. A process has a virtual address space, executable code, open handles to system objects, a security context, a unique process identifier, environment variables, a priority class, minimum and maximum working set sizes, and at least one thread of execution. Each process is started with a single thread, often called the primary thread, but can create additional threads from any of its threads. A thread is the entity within a process that can be scheduled for execution. All threads of a process share its virtual address space and system resources. In addition, each thread maintains exception handlers, a scheduling priority, thread local storage, a unique thread identifier, and a set of structures the system will use to save the thread context until it is scheduled. On a multiprocessor computer, the system can simultaneously execute as many threads as there are processors on the computer. A job object allows groups of processes to be managed as a unit. Job objects are namable, securable, sharable objects that control attributes of the processes associated with them. Operations performed on the job object affect all processes associated with the job object. Examples include enforcing limits such as working set size and process priority or terminating all processes associated with a job. A thread pool is a collection of worker threads that efficiently execute asynchronous callbacks on behalf of the application. The thread pool is primarily used to reduce the number of application threads and provide management of the worker threads. A fiber is a unit of execution that must be manually scheduled by the application. Fibers run in the context of the threads that schedule them. Each thread can schedule multiple fibers. In general, fibers do not provide advantages over a welldesigned multithreaded application. However, using fibers can make it easier to port applications that were designed to schedule their own threads. From a system standpoint, a fiber assumes the identity of the thread that runs it. For example if a fiber accesses thread local storage, it is accessing the thread local storage of the thread that is running it. In addition, if a fiber calls the ExitThread function, the thread that is running it exits. However, a fiber does not have all the same state information associated with it as that associated with a thread. The only state information maintained for a fiber is its stack, a subset of its registers, and the fiber data provided during fiber creation. The saved registers are the set of registers typically preserved across 196 Chapter 4 / Threads a function call. Fibers are not preemptively scheduled. A thread schedules a fiber by switching to it from another fiber. The system still schedules threads to run. When a thread that is running fibers is preempted, its currently running fiber is preempted but remains selected. User-mode scheduling (UMS) is a lightweight mechanism that applications can use to schedule their own threads. An application can switch between UMS threads in user mode without involving the system scheduler, and regain control of the processor if a UMS thread blocks in the kernel. Each UMS thread has its own thread context instead of sharing the thread context of a single thread. The ability to switch between threads in user mode makes UMS more efficient than thread pools for short-duration work items that require few system calls. UMS is useful for applications with high performance requirements that need to efficiently run many threads concurrently on multiprocessor or multicore systems. To take advantage of UMS, an application must implement a scheduler component that manages the application’s UMS threads and determines when they should run. Management of Background Tasks and Application Lifecycles Beginning with Windows 8, and carrying through to Windows 10, developers are responsible for managing the state of their individual applications. Previous versions of Windows always give the user full control of the lifetime of a process. In the classic desktop environment, a user is responsible for closing an application. A dialog box might prompt them to save their work. In the new Metro interface, Windows takes over the process lifecycle of an application. Although a limited number of applications can run alongside the main app in the Metro UI using the SnapView functionality, only one Store application can run at one time. This is a direct consequence of the new design. Windows Live Tiles give the appearance of applications constantly running on the system. In reality, they receive push notifications and do not use system resources to display the dynamic content offered. The foreground application in the Metro interface has access to all of the processor, network, and disk resources available to the user. All other apps are suspended and have no access to these resources. When an app enters a suspended mode, an event should be triggered to store the state of the user’s information. This is the responsibility of the application developer. For a variety of reasons, whether it needs resources or because an application timed out, Windows may terminate a background app. This is a significant departure from the Windows operating systems that precede it. The app needs to retain any data the user entered, settings they changed, and so on. That means you need to save your app’s state when it’s suspended, in case Windows terminates it, so you can restore its state later. When the app returns to the foreground, another event is triggered to obtain the user state from memory. No event fires to indicate termination of a background app. Rather, the application data will remain resident on the system, as though it is suspended, until the app is launched again. Users expect to find the app as they left it, whether it was suspended or terminated by Windows, or closed by the user. Application developers can use code to determine whether it should restore a saved state. 4.4 / WINDOWS PROCESS AND THREAD MANAGEMENT 197 Some applications, such as news feeds, may look at the date stamp associated with the previous execution of the app and elect to discard the data in favor of newly obtained information. This is a determination made by the developer, not by the operating system. If the user closes an app, unsaved data is not saved. With foreground tasks occupying all of the system resources, starvation of background apps is a reality in Windows. This makes the application development relating to the state changes critical to the success of a Windows app. To process the needs of background tasks, a background task API is built to allow apps to perform small tasks while not in the foreground. In this restricted environment, apps may receive push notifications from a server or a user may receive a phone call. Push notifications are template XML strings. They are managed through a cloud service known as the Windows Notification Service (WNS). The service will occasionally push updates to the user’s background apps. The API will queue those requests and process them when it receives enough processor resources. Background tasks are severely limited in the usage of processor, receiving only one processor second per processor hour. This ensures that critical tasks receive guaranteed application resource quotas. It does not, however, guarantee a background app will ever run. The Windows Process Important characteristics of Windows processes are the following: Windows processes are implemented as objects. A process can be created as a new process or as a copy of an existing process. An executable process may contain one or more threads. Both process and thread objects have built-in synchronization capabilities. Figure 4.10, based on one in [RUSS11], illustrates the way in which a process relates to the resources it controls or uses. Each process is assigned a security access token, called the primary token of the process. When a user first logs on, Windows creates an access token that includes the security ID for the user. Every process that is created by or runs on behalf of this user has a copy of this access token. Windows uses the token to validate the user’s ability to access secured objects, or to perform restricted functions on the system and on secured objects. The access token controls whether the process can change its own attributes. In this case, the process does not have a handle opened to its access token. If the process attempts to open such a handle, the security system determines whether this is permitted, and therefore whether the process may change its own attributes. Also related to the process is a series of blocks that define the virtual address space currently assigned to this process. The process cannot directly modify these structures, but must rely on the virtual memory manager, which provides a memoryallocation service for the process. Finally, the process includes an object table, with handles to other objects known to this process. Figure 4.10 shows a single thread. In addition, the process has access to a file object and to a section object that defines a section of shared memory. 198 Chapter 4 / Threads Process and Thread Objects The object-oriented structure of Windows facilitates the development of a generalpurpose process facility. Windows makes use of two types of process-related objects: processes and threads. A process is an entity corresponding to a user job or application that owns resources, such as memory and open files. A thread is a dispatchable unit of work that executes sequentially and is interruptible, so the processor can turn to another thread. Each Windows process is represented by an object. Each process object includes a number of attributes and encapsulates a number of actions, or services, that it may perform. A process will perform a service when called upon through a set of published interface methods. When Windows creates a new process, it uses the object class, or type, defined for the Windows process as a template to generate a new object instance. At the time of creation, attribute values are assigned. Table 4.3 gives a brief definition of each of the object attributes for a process object. A Windows process must contain at least one thread to execute. That thread may then create other threads. In a multiprocessor system, multiple threads from the same process may execute in parallel. Table 4.4 defines the thread object attributes. Note some of the attributes of a thread resemble those of a process. In those cases, the thread attribute value is derived from the process attribute value. For example, the thread processor affinity is the set of processors in a multiprocessor system that may execute this thread; this set is equal to or a subset of the process processor affinity. Figure 4.10 A Windows Process and Its Resources Process object Access token Virtual address descriptors Thread x File y Section z Handle1 Handle2 Handle3 Available Handle table objects 4.4 / WINDOWS PROCESS AND THREAD MANAGEMENT 199 Note one of the attributes of a thread object is context, which contains the values of the processor registers when the thread last ran. This information enables threads to be suspended and resumed. Furthermore, it is possible to alter the behavior of a thread by altering its context while it is suspended. Process ID A unique value that identifies the process to the operating system. Security descriptor Describes who created an object, who can gain access to or use the object, and who is denied access to the object. Base priority A baseline execution priority for the process’s threads. Default processor affinity The default set of processors on which the process’s threads can run. Quota limits The maximum amount of paged and nonpaged system memory, paging file space, and processor time a user’s processes can use. Execution time The total amount of time all threads in the process have executed. I/O counters Variables that record the number and type of I/O operations that the process’s threads have performed. VM operation counters Variables that record the number and types of virtual memory operations that the process’s threads have performed. Exception/debugging ports Interprocess communication channels to which the process manager sends a message when one of the process’s threads causes an exception. Normally, these are connected to environment subsystem and debugger processes, respectively. Exit status The reason for a process’s termination. Table 4.3 Windows Process Object Attributes Thread ID A unique value that identifies a thread when it calls a server. Thread context The set of register values and other volatile data that defines the execution state of a thread. Dynamic priority The thread’s execution priority at any given moment. Base priority The lower limit of the thread’s dynamic priority. Thread processor affinity The set of processors on which the thread can run, which is a subset or all of the processor affinity of the thread’s process. Thread execution time The cumulative amount of time a thread has executed in user mode and in kernel mode. Alert status A flag that indicates whether a waiting thread may execute an asynchronous procedure call. Suspension count The number of times the thread’s execution has been suspended without being resumed. Impersonation token A temporary access token allowing a thread to perform operations on behalf of another process (used by subsystems). Termination port An interprocess communication channel to which the process manager sends a message when the thread terminates (used by subsystems). Thread exit status The reason for a thread’s termination. Table 4.4 Windows Thread Object Attributes 200 Chapter 4 / Threads Multithreading Windows supports concurrency among processes because threads in different processes may execute concurrently (appear to run at the same time). Moreover, multiple threads within the same process may be allocated to separate processors and execute simultaneously (actually run at the same time). A multithreaded process achieves concurrency without the overhead of using multiple processes. Threads within the same process can exchange information through their common address space and have access to the shared resources of the process. Threads in different processes can exchange information through shared memory that has been set up between the two processes. An object-oriented multithreaded process is an efficient means of implementing a server application. For example, one server process can service a number of clients concurrently. Thread States An existing Windows thread is in one of six states (see Figure 4.11): 1. Ready: A ready thread may be scheduled for execution. The Kernel dispatcher keeps track of all ready threads and schedules them in priority order. 2. Standby: A standby thread has been selected to run next on a particular processor. The thread waits in this state until that processor is made available. If the standby thread’s priority is high enough, the running thread on that processor may be preempted in favor of the standby thread. Otherwise, the standby thread waits until the running thread blocks or exhausts its time slice. Figure 4.11 Windows Thread States Runnable Not runnable Pick to run Switch Preempted Block/ suspend Unblock/resume Resource available Resource available Unblock Resource not available Terminate Standby Ready Running Transition Waiting Terminated 4.4 / WINDOWS PROCESS AND THREAD MANAGEMENT 201 3. Running: Once the Kernel dispatcher performs a thread switch, the standby thread enters the Running state and begins execution and continues execution until it is preempted by a higher-priority thread, exhausts its time slice, blocks, or terminates. In the first two cases, it goes back to the Ready state. 4. Waiting: A thread enters the Waiting state when (1) it is blocked on an event (e.g., I/O), (2) it voluntarily waits for synchronization purposes, or (3) an environment subsystem directs the thread to suspend itself. When the waiting condition is satisfied, the thread moves to the Ready state if all of its resources are available. 5. Transition: A thread enters this state after waiting if it is ready to run, but the resources are not available. For example, the thread’s stack may be paged out of memory. When the resources are available, the thread goes to the Ready state. 6. Terminated: A thread can be terminated by itself, by another thread, or when its parent process terminates. Once housekeeping chores are completed, the thread is removed from the system, or it may be retained by the Executive6 for future reinitialization. Support for OS Subsystems The general-purpose process and thread facility must support the particular process and thread structures of the various OS environments. It is the responsibility of each OS subsystem to exploit the Windows process and thread features to emulate the process and thread facilities of its corresponding OS. This area of process/thread management is complicated, and we give only a brief overview here. Process creation begins with a request for a new process from an application. The application issues a create-process request to the corresponding protected subsystem, which passes the request to the Executive. The Executive creates a process object and returns a handle for that object to the subsystem. When Windows creates a process, it does not automatically create a thread. In the case of Win32, a new process must always be created with an initial thread. Therefore, the Win32 subsystem calls the Windows process manager again to create a thread for the new process, receiving a thread handle back from Windows. The appropriate thread and process information are then returned to the application. In the case of POSIX, threads are not supported. Therefore, the POSIX subsystem obtains a thread for the new process from Windows so that the process may be activated but returns only process information to the application. The fact that the POSIX process is implemented using both a process and a thread from the Windows Executive is not visible to the application. When a new process is created by the Executive, the new process inherits many of its attributes from the creating process. However, in the Win32 environment, this process creation is done indirectly. An application client process issues its process creation request to the Win32 subsystem; then the subsystem in turn issues a process request to the Windows executive. Because the desired effect is that the new process inherits characteristics of the client process and not of the server process, Windows enables the 6The Windows Executive is described in Chapter 2. It contains the base operating system services, such as memory management, process and thread management, security, I/O, and interprocess communication. 202 Chapter 4 / Threads subsystem to specify the parent of the new process. The new process then inherits the parent’s access token, quota limits, base priority, and default processor affinity.