Return to lecture notes index
January 17, 2013 (Lecture 2)

The Area We Call "Systems" -- And the Funny Creatures We Call "Systems People"

When I'm hanging out in the "Real world", people often ask me about my job. I usually explain that I am a teacher. Everyone understands what a teacher does. We talk for a living. Beyond that, I'm safe. Everyone knows, "Those who can, do. Those who can't, teach."

When people ask me what I teach, I tell them, "Computer Science". Oddly enough, they only hear the first word, "Computer". Sorry, ya'll, I don't do windows. You'll need IT for that. This brings me to two questions, "What is the area we call, Comptuer Systems?" and, "How does Distributed Systems fit in?"

When I explain my area of interest to every day folks, I like to tell them that in "Systems" we view the computing landscape as commerce from the perspective of the air traffic system or the system of highways and roadways. There is a bunch of work that needs to get done, a bunch of resources that need to be used to get it done, and a whole lot of management to make it work.

And, like our view of commerce, it only gets interesting when it scales to reach scarcity and when bad things happen. The world can't be accurately descibed in terms of driveways, lemonaide stands, and sunny days. Instead, we care about how our roadways and airways perform during rush hour, in the rain, when there is a big game, and, by the way, bad things happen to otherwise good drivers along the way. In otherwords, our problem space is characterized by scarcity, failure, interaction, and scale.

Distributed Systems, In Particular

"Systems people" come in all shapes and sizes. They are interested in such problems as operating systems, networks, databases, and distributed systems. This semester, we are focusing mostly on "Distributed systems", though we'll touch on some areas of networks, and monolithic databases and operating systems.

Distributed systems occur when the execution of user work involved managing state which is connected somewhat weakly. In other words, distributed systems generally involve organizing resources connected via a network that has more latency, less bandwidth, and/or a higher error rate than can be safely ignored.

This is a different class of problems, for example, than when the limiting factors might include processing, storage, memory, or other units of work. There is tremendous complexity in scheduling process to make efficient use of scarce processors, managing virtual memory, or processing information from large attached data stores, as might occur in monolithic operating systems or databases. It is also a different class of problems than managing the fabric, itself, as is the case with networks.

Exploring the Model

When I've taught Operating Systems, I've begun with a picture that looks like the one below. If you didn't take OS, please don't worry -- everything on the picture, almost, should be familiar to you. It contains the insides of a computer: memory and memory controllers, storage devices and their controllers, processors, and the bus that ties them all together.

This time however, the bus isn't magical. It isn't a fast, reliable, predictable communication channel called that always works and maintains a low latency and high bandwidth. Instead, it is a simple, cheap, far-reaching commodity network that may become slow and bogged down and/or lose things outright. It might become partitions. And, it might not deliver messages in the same order that they were sent.

To reinforce the idea that this is a commodity network, like the Internet, I added a few PDAs to the picture this time. Remember, the network isn't necessarily wired -- and all of the components aren't necessarily of the same type.

Furthermore, there is no global clock or hardware support for synchronization. And, to make things worse, thr processors aren't necessarily reliable, and nor is the RAM or anything else. For those that are familiar with them, snoopy caches aren't practical, either.

In other words, all of the components are independent, unreliable devices connected by an unreliable, slow, narrow, and disorganized network.

What's the Good News?

The bottom line is that, despite the failure, uncertainty, and lack of specialized hardware support, we can build and effectively use systems that are an order of magnitude more powerful. In fact we can do this while providing a more available, more robust, more convenient solution. This semester, we'll learn how.

Distributed Systems vs. Parallel Systems

Often we hear the terms "Distributed System" and "Parallel System." What is the difference?

Not a whole lot and a tremendous amount -- all at the same time. "Distributed System" often refers to a systems that is to be used by multiple (distributed) users. "Parallel System" often has the connotation of a system that is designed to have only a single user or user process. Along the same lines, we often hear about "Parallel Systems" for scientific applications, but "Distributed Systems" in e-commerce or business applications.

"Distributed Systems" generally refer to a cooperative work environment, whereas "Parallel Systems" typically refer to an environment designed to provide the maximum parallelization and speed-up for a single task. But from a technology perspective, there is very little distinction.

Does that suggest that they are the same? Well, not exactly. There are some differences. Security, for example, is much more of a concern in "Distributed Systems" than in "Parallel Systems". If the only goal of a super computer is to rapidly solve a complex task, it can be locked in a secure facility, physically and logically inaccessible -- security problem solved. This is not an option, for example, in the design of a distributed database for e-commerce. By its very nature, this system must be accessible to the real world -- and as a consequence must be designed with security in mind.


We'll hear about many abstractions this semester -- we'll spend a great deal of time discussing various abstractions and how to model them in software. So what is an abstraction?

An abstraction is a representation of something that incorporates the essential or relevent properties, while neglecting the irrelevant details. Throughout this semester, we'll often consider something that exists in the real world and then distill it to those properties that areof concern to us. We'll often then take those properties and represent them as data structures and algorithms that that represent the "real world" items within our software systems.

The Task

The first abstraction that we'll consider is arguably the most important -- a represention of the work that the system will do on behalf of a user (or, perhpas, itself). I've used a lot of different words to describe this so far: task, job, process, &c. But I've never been very specific about what I've meant -- to be honest, I've been a bit sloppy.

This abstraction is typically called a task. In a slightly different form, it is known as a process. We'll discuss the subtle difference when we discuss threads. The short version of the difference is that a task is an abstraction that represents the instance of a program in execution, whereas a process is a particular type of task with only one thread of control. But, for now, let's not worry about the difference.

If we say that a task is an instance of a program in execution, what do we mean? What is an instance? What is a program? What do we mean by execution?

A program is a specification. It contains defintions of what type fo data is stored, how it can be accessed, and a set of instructions that tells the computer how to accomplish something useful. If we think of the program as a specification, much like a C++ class, we can think of the task as an instance of that class -- much like an object built from the specification provided by the program.

So, what do we mean by "in execution?" We mean that the task is a real "object" not a "class." Most importantly, the task has state associated with it -- it is in the process of doing something or changing somehow. Hundreds of tasks may be instances of the same program, yet they might behave very differently. This happens because the tasks were exposed to different stimuli and their changed accordingly.

Representing a Task in Software

How do we represent a task within the context of an operating system? We build a data structure, sometimes known as a task_struct or (for processes) a Process Control Block (PCB) that contains all of the information our OS needs about the state of the task. This includes, among many other things:

When a context switch occurs, it is this information that needs to be saved and restored to change the executing process.

The Life Cycle of a Process (For our present purposes, a Task)

Okay. So. Now that we've got a better understanding of the role of the OS and the nature of a system call, let's move in the direction of this week's lab. It involves the management of processes. So, let's begin that discussion by considering the lifecycle of a process:

A newly created process is said to be ready or runnable. It has everything it needs to run, but until the operating system's schedule dispatches it onto a processor, it is just waiting. So, it is put onto a list of runnable processes. Eventually, the OS selects it, places it onto the processor, and it is actually running.

If the timer interrupts its execution and the OS decides that it is time for another process to run, the other process is said to preempt it. The preempted process returns to the ready/runnable list until it gets the opportunity to run again.

Sometimes, a running process asks the operating system to do something that can take a long time, such as read from the disk or the network. When that happens, the operating system doesn't want to force the processor to idle while the process is waiting for the slow action. Instead, it blocks the process. It moves the proces to a wait list associated with the slow resource. It then chooses another process from the ready/runnabel list to run.

Eventually, the resource, via an interrupt, will let the OS know that the process can again be made ready to run. The OS will do what it needs to do, and ready, a.k.a., make runnable, the previously blocked process by moving it to the ready/runnable list.

Eventually a program may die. It might call exit under the programmer's control, in which case it is said to exit or it might end via some exception, in which case the more general term, terminated might be more descriptive. When this happens, the process doesn't immediately go away. Instead, it is said to be a zombie. The process remains a zombie until its parent uses wait() or waitpid() to collect its status -- and set it free.

If the parent died before the child, or if it died before waiting for the child, the child becomes an orphan. Shoudl this happen, the OS will reparent the orphan process to a special process called init. Init, by convention, has pid 1, and is used at boot time to start up other processes. But, it also has the special role of waiting() for all of the orphans that are reparented to it. In this way, all processes can eventually be cleaned up. When a process is set free by a wait()/waitpid(), it is said to be reaped.

Tasks in Distributed Systems

In distributed systems, we find that the various resources needed to perform a task are scattered across a network. This blurs the distinction between a process and a task and, for that matter, a task and a thread. In the context of distributed systems, a process and a thread are interchangable terms -- they represent something that the user wants done.

But, task has an interesting and slightly nuianced meaning. A task is the collection of resources configured to solve a particular problem. A task contains not only the open files and communication channels -- but also the threads (a.k.a. processes). Distributed Systems people see a task as the enviornment in which work is done -- and the thread (a.k.a. process) as the instance of that work, in progress.

I like to explain that a task is a factory -- all of the means of production scattered across many assembly lines. The task contains the machinery and the supplies -- as well the processes that are ongoing and making use of them.

Creating New Tasks

One of the functions of the operating system is to provide a mechanism for existing tasks to create new tasks. When this happens, we call the original task the parent. The new task is called the child. It is possible for one task to have many children. In fact, even the children can have children.

In UNIX, child tasks can either share resources with the parent or obtain new resources. But existing resources are not partitioned.

In UNIX when a new task is created, the child is a clone of the parent. The new task can either continue to execute with a copy of the parent image, or load another image. Well talk more about this soon, when we talk about the fork() and exec-family() of calls.

After a new task is created, the parent may either wait()/waitpid() for the child to finish or continue and execute concurrently (real or imaginary) with the child.

Fork -- A traditional implementation

fork() is the system call that is used to create a new task on UNIX systems. In a traditional implementation, it creates a new task by making a nearly exact copy of the parent. Why nearly exact? Some things don't make sense to be duplicated exactly, the ID number, for example.

The fork() call returns the ID of the child process in the parent and 0 in the child. Other than this type of subtle differences, the two tasks are very much alike. Execution picks up at the same point in both.

If execution picks up at the same point in both, how can fork() return something different in each? The answer is very straightforward. The stack is duplicated and a different value is placed on top of each. (If you don't remeber what the stack is, don't worry, we'll talk about it soon -- just realize that the return value is different).

The difference in the return value of the fork() is very significant. Most programmers check the result of the fork in order to determine whether they are currently the child or parent. Very often the child and parent to very different things.

The Exec-family() of calls

Since the child will often serve a very different purpose that its parent, it is often useful to replace the child's memory space, that was cloned form the parent, with that of another program. By replace, I am referring to the following process:
  1. Deallocate the process' memory space (memory pages, stack, etc).
  2. Allocate new resources
  3. Fill these resources with the state of a new process.
  4. (Some of the parent's state is preserved, the group id, interrupt mask, and a few other items.)
Fork w/copy-on-write
Copying all of the pages of memory associated with a process is a very expensive thing to do. It is even more expensive considering that very often the first act of the child is to deallocate this recently created space.

One alternative to a traditional fork implementation is called copy-on-write. the details of this mechanism won't be completely clear until we study memory management, but we can get the flavor now.

The basic idea is that we mark all of the parent's memory pages as read-only, instead of duplicating them. If either the parent or any child try to write to one of these read-only pages, a page-fault occurs. At this point, a new copy of the page is created for the writing process. This adds some overhead to page accesses, but saves us the cost of unnecessarly copying pages.


Another alternative is also available -- vfork(). vfork is even faster, but can also be dangerous in the worng hands. With vfork(), we do not duplicate or mark the parent's pages, we simply loan them, and the stack frame to the child process. During this time, the parent remains blocked (it can't use the pages). The dangerous part is this: any changes the child makes will be seen by the aprent process.

vfork() is most useful when it is immediately followed by an exec_(). This is because an exec() will create a completely new process-space, anyway. There is no reason to create a new task space for the child, just to have it throw it away as part of an exec(). Instead, we can loan it the parent's space long enough for it to get started (exec'd).

Although there are several (4) different functions in the exec-family, the only difference is the way they are parameterizes; under-the-hood, they all work identically (and are often one).

After a new task is created, the parent will often want to wait for it (and any siblings) to finish. We discussed the defunct and zombie states last class. The wait-family of calls is used for this purpose.

A Process's Memory

Before we talk about how a tasks' memory is laid out, let's first look at the simpler case of a process -- a task with only one thread of control.

Please note that the heap grows upward through dynamic allocation (like malloc) and the stack grows downward as stack frames are added throguh function calls. Such things as return addresses, return values, parameters, local variables, and other state are stored in the runtime stack.

The Conceptual Thread

In our discussion of tasks we said that a task is an operating system abstraction that represents the state of a program in execution. We learned that this state included such things as the registers, the stack, the memory, and the program counter, as well as software state such as "running," "blocked", &c. We also said that the processes on a system compete for the systems resources, especially the CPU(s).

Another operating system abstraction is called the thread. A thread, like a task, represents a discrete piece of work-in-progress. But unlike tasks, threads cooperate in their use of resources and in fact share many of them.

We can think of a thread as a task within a task. Among other things threads introduce concurrency into our programs -- many threads of control may exist. Older operating systems didn't support threads. Instead of tasks, they represented work with an abstraction known as a process. The name process, e.g. first do ___, then do ____, if x then do ____, finally do ____, suggests only one thread of control. The name task, suggests a more general abstraction. For historical reasons, colloquially we often say process when we really mean task. From this point forward I'll often say process when I mean task -- I'll draw our attention to the difference, if it is important.

The Implementation of Threads

All of the threads within a process exist within the context of that process. They share the code section, data section, and operating system resources such as open files.

But they do not share all resources. Since each thread executes independently, each thread has its own understanding of the stack and of the registers.

The good part about sharing so many resources is that switching execution among threads is less expensive than it is among processes.

The bad part is that unlike the protection that exists among processes, the operating system can not prevent threads from interfering with each other -- they share the same process space.

Kernel Threads

The most primitive implementations of threads were invisible to the user. They existed only within the kernel. Many of the kernel daemons, such at the page daemon, were implemented as threads. Implementing different operating system functions as threads made sense for many reasons:

User Threads

But the UNIX developers couldn't keep such a great thing as their own private secret for long. Users began to use threads via thread libraries. Thread libraries gave users the illusion of threads, without any involvement form the kernel. When a process containing threads is executed, the thread scheduler, within the process, is run. This scheduler selects which thread should run and for how long. If a thread should block, the scheduler can select to run another thread within the same process.

This implementation of threads is actually much more than an illusion. It gives users the ability to write very efficient programs. These programs can switch among threads and share resurces with very little overhead. To switch threads, the registers must be saved and restored and the stack must be switched. No expensive context switch is required.

Another advantage is that user level threads are implemented entirely by a thread library -- from the interface to the scheduling. The kernel doesn't see them or know about them.

Kernel-Supported User-Level Threads

Kernel threads are great for kernel writers and user threads answer many of the needs of users, but they are not perfect. Consider these examples:

To address these needs, we need to have a kernel supported user thread. That is to say, we need a facility for threads to share resources within a process, but we also need the ability of the kernel to preempt, schedule, and dispatch threads. This type of thread is called a kernel supported user thread or a light-weight process (LWP). A light-weight process is in contrast with a heavy-weight process otherwise known as a process or task.

Our model of the universe has gone from looking like this:

To looking like this:

A More Complex Model

In some models, such as that used by Solaris, it is also possible to assign several kernel-supported threads to a single process without assigning them to specific user-level threads. In this case, the process will have more opportunities to be seen by the OS's CPU scheduler. On multiprocessor systems, the maximum level of concurrency is determined by the number of LWPs assigned to the process (of course this is further limited by the number of threads that are runnable within the process and the number of available CPUs).

In the context of Solaris, an LWP is a user-visible kernel thread. In some ways, it might be better to view a Solaris LWP as a virtual light weight processor (this is Kesden nomenclature!). This is because pools of LWPs can be assigned to the same task. Threads within that task are then scheduled to run on available LWPs, much like processes are scheduled to run on available processors.

In truth, LWPs are anything but light weight. They are lighter weight than (heavy weight) processes -- but they require far more overhead than user-level threads without kernel support. Context-switching among user-level threads within a process is much, much cheaper than context switching among LWPs. But, switching among LWPs can lead to greater concurrency for a task when user-level threads block within the kernel (as opposed to within the process such that the thread scheduler can run another).

The diagram below shows LWP's associated with tasks and kernel threads, as well as kernel threads without an associated LWP and several different associations between user-level threads and the LWP(s) assigned to the process.

LWP's offer a convenient and flexible compromise between user-threads and separate processes. But it is important to realize that they are bulky structures:

Another Model

You might observe that threads within the same task share many of the same resources -- the most significant difference is that they have different stacks. This is exactly how Linux implements threads -- by leveraging the machinery it uses to create new processes.

In Linux, the system call beneath a fork() is known as clone(). It does everything that fork() does -- but wiht a lot more flexibility. It allows a newly created child to be created from the template of its parent by either copying, sharing, or recreating various resources from the parent. So, you can imagine that if all of the important resources are copies -- clone() is, in effect, a fork().

But, imagine two different processes that, in essence, share everything but the stack. These processes are sharing global memory, so they are really in the smae memory context -- but they have differnt stacks, so can be doing different things. They are, in effect, threads.

So, as Linux evolved, they created kernel-supported threads by leveraging clone() to create new processes that shared resources. They then created a new task abstraction that aggregated related processes. So, in the Linux model, within the kernel, a task is genuinely a collection of processes. From the user's perspective, these thread-like processes are presented as threads by the user-level thread library.