Scheduling the access of processes to non-sharable resources is a fundamental part of an operating system's job. The same is true of the thread scheduler withina user-level thread library. The CPU is the most important among these resource, because it typically has the highest contention. The high cost of additional CPUs, both in terms of the price of the processor and the cost of the technology in massively parallel systems, ensures that most systems don't have a sufficient surpless of CPUs to allow for their wasteful use.
The primary objective of CPU scheduling is to ensure that as many jobs are running at a time as is possible. On a single-CPU system, the goal is to keep one job running at all times.
Multiprogramming allows us to keep many jobs ready to run at all times. Although we can not concurrently run more jobs than we have available processors, we can allow each processor to be running one job, while other jobs are waiting for I/O or other events.
Observation: The CPU-I/O Burst Cycle
During our discussion of scheduling, I may make reference to the CPU-I/O burst cycle. This is a reference to the observation that programs usually have a burst of I/O (when the collect data) followed by a burst of CPU (when they process it and request more). These bursts form a cycle of execution.
Some processes have long bursts of CPU usage, followed by short bursts of I/O. We say that thee jobs are CPU Bound.
Some processes have long bursts of I/O, followed by short bursts of CPU. We say that thee jobs are I/O Bound.
The CPU Scheduler
Some schedulers are only invoked after a job finishes executing or voluntarily yield the CPU. This type of scheduler is called a non-premptive scheduler.
- A process blocks itself waiting for a resource or event
- A process terminates
But, in order to support interaction, most modern schedulers are premeptive. They make use of a hardware timer to interrupt running jobs. When a scheduler hardware interrupt occurs, the scheduler's ISR is invoked and it runs. When this happens, it can continue the previous task or run another. Before starting a new job, the scheduler sets the hardware timer to generate an interrupt after a particular amount of time. This time is known as the time quantum. It is the amount of time that a job can run without interruption.
A preemptive scheduler may be invoked under the following four circumstances:
- A process blocks itself waiting for a resource or event
- A process terminates
- A process moves from running to ready (interrupt)
- A process moves from waiting to ready (blocking condition satisfied)
Once the CPU scheduler selects a process for execution, it is the job of the dispatcher to start the selected process. Starting this process involves three steps:Scheduling Algorithms
- Switching context
- Switching to user mode
- Jumping to the proper location in the program to start or resume execution
The latency introduced by the dispatch is called the dispatch latency. Obviously, this should be as small as possible -- but is most critical in real-time system, those systems that must meet deadlines associated with real world events. These systems are often associted with manufacturing systems, monitoring systems, &c. Actually, admission is also much more important in these systems -- a job isn't automically admitted, it is only admitted if the system can verify that enough resources are available to meet the deadlines associated with the job. But real-time systems are a different story -- back to today's tale.
So, given a collection of tasks, how might the OS (or the user-level thread scheduler) decide which to place on an available CPU?
First Come, First Serve (FCFS)
First In, First Out (FIFO)
FCFS is the simplest algorithm. It should make sense to anyone who has waited in line at the deli, bank, or check-out line, or to anyone who has ever called a customer service telephone number, "Your call will be answered in the order in which it was received."
The approach is very simple. When a job is submitted, it enters the ready queue. The oldest job (has been in the ready queue for the longest time) in the ready queue is always selected to be dispatched. The algorithm is non-premptive, so the job will run until it voluntarily gives up the CPU by blocking or terminating. After a blocked process is satisfied and returns to the ready queue, it enters at the end of the line.
This algorithm is very easy to implement, and it is also very fair and consequently starvation-free. No characteristic of a job bias its placement in the queue. But it does have several disadvantages.
- Because short jobs can be scheduled after very long jobs, the average wait time can be increased. Unlike the grocery store, the algorithm has no "Express Lane."
- Since the algorithm is non-premptive, there can be problems on interactive systems. A long job may prevent responsiveness to the user.
- A student actually pointed out this more subtle property during class in the form of a question (good job!) It is possible for one CPU bound process to hog the CPU for a long time. In the meantime several I/O bound processes, that could execute concurrently with the CPU-bound process, can't start.
I didn't give this phenomenon a name in class, but I probably should have -- it is called the convoy effect, because cycle after cycle, the I/O bound processes will "follow" the CPU bound processes, without overlapping.
Another apporach is to consider the expected length of each processes's next CPU burst and to run the process with the shortest burst next. This algorithm optimizes the average waiting time of the processes. This is because moving a shorter job ahead of a longer job helps the shorter job more than it hurts the longer job. Recall my lunchroom story -- those who ate early had no lines, although there wasn't anyone to vouch for this.
Unfortunately, we have no good way of knowing for sure the length of any jobs next CPU burst. In practice this can be estimated using an exponential average of the jobs recent CPU usage. In the past, programmers estimated it -- but if their jobs went over their estimate, they were killed. programmers got very good at "The Price Is Right."
Sometimes this algorithm is premeptive. A job can be prempted if another job arrives that has a shorter execution time. This is flavor is often called shortest-remaining-time-first (SRTF).
Shortest CPU time to Completion First (STCF). The process that will complete first runs whenever possible. The other processes only run when the first process is busy with an I/O event.
But, much like SJF, the CPU time is not known in advance.
Priority Scheduling (PRI)
Priority scheduling is designed to strictly enforce the goals of a system. Important jobs always run before less important jobs. If this scheduling discipline is implemented preemeptively, more important jobs will preempt less important jobs, even if they are currently running.
The bigest problem with priority-based scheduling is starvation. it is possible that low priority jobs will never execute, if more important jobs continually arrive.
Round Robin Scheduling (RR)
Round Robin scheduling can be thought of as a preemptive version of FCFS. Jobs are processed in a FCFS order from the run queue. As with FCFS, if they block, the next process can be scheduled. And when a blocked process returns to the ready queue, it is placed at the end of the list.
The difference is that each process is given a time quantum or time slice. A hardware timer interrupt preemept the running process, if it is still running after this fixed amount of time. The scheduler can then dispatch the next process in the queue.
With an appropriate time quantum, this process offeres a better average case performance than FCFS without the guesswork of SJF.
If the time quantum is very, very small, an interesting effect is produced. If there are N processes, each process executes as if it were running on its own private CPU running at 1/N th the speed. This effect is called processor sharing.
If the time quantum is very, very large -- large enough that the processes generally complete before it expires, this approach approximates FCFS.
The time quantum can be selected to balance the two effects.
Everything in this section goes above and beyond lecture. It is just here for those who happen to be curious for more real-world detail.
Multilevel Queue Scheduling (MQS)
MQS is similar to PRI, except that the jobs arrive sorted by their priority. For example, all system jobs may have a higher priority than interactive jobs, which enjoy a higher priority than batch jobs. Jobs of different priorities are placed into different queues.
In some implementations, jobs in all higher priority queues must be executed before jobs in any lower priority queue. This absolute approach can lead to starvation in the same way as its simplier cousin, PRI. In some preemptive implementations, a lower-priority process will be returned to its ready queue, if a higher-priority process arrives.
Another approach is to time-slice among the queues. Higher priority queus can be given longer or more frequent time slices. This approach prevents absolute starvation.
Multilevel Feedback Queue (MFQ)
In the multilevel queuing system we described above, there is no social mobility. When aprocess arrives, it is placed in a queue based on its initial classification. It remains in this queue throughout its lifetime.
In a MFQ system, a process's priority can change as the result of its CPU utilization. Processes that have longer CPU bursts are lowered in importance. Processes that are I/O bound and frequently release the CPU prematurely by blocking are increased in importance.
This system prevents starvation and gives I/O bound processes the change to be dispatched and overlap with CPU bound processes. It fights the convoy effect.
Scheduling among queues can be done by varying the lenght or frequence of the time slice. Scheduling within each queue can also be varied.
Although this system sounds very pretty, it is very complex. In general this type of system is defined by the following parameters:
- the number of queues
- the method of scheduling for each queue
- the method of scheudling among the queues
- the method of promoting a process
- the method of demoting a process
Traditional UNIX Scheduling - Introduction
I thought it would be interesting to spend some time considering scheduling in a real-world operating system. Today we'll talk about what I call "traditional" UNIX scheduling. This scheduling system was used, with little variation through AT&T SVR3 and 4.3BSD. Newer UNIX's use more sophisticated approaches, but this is a good place to start.
The scheduling in these systems was priority based. The priority of a process ranged from 0 - 127. Counterintuitively, lower priorities represent more important processes.
The range of priorities is partitioned. Priorities in the range of 0 - 49 are reserved for system processes. Both user and system processes may have priorities over the full range from 0 - 127. Ths prevents user processes from interfereing with more important system tasks.
A processes ultimate scheduling priority varies with several factors. The accounting of these factors is kept in the proc structure, which contains the following fields:
- p_pri - current scheduling priority
- p_usrpri - the process's priority in user mode
- p_cpu - a measure of the prcoess's recent CPU usage
- p_nice - a user supplied measure of the process's importance.
Elevated Priority in System Mode
p_pri is a number in the range of 0 - 127 that represents the priority of the process. This is the value of that the scheduler considers when selecting a process to be dispatched. This value is normally the same as p_usrpri. It however, may be lowered (making the process more important) while the process is making a system call.
Traditional UNIX systems did not have preemptive kernels. This meant that only one process could be in the kernel at a time. If a process blocked while in a system call, other user processes could run, but not other system calls or functions. For this reason a process which had blocked while in a system call often would have its p_pri value lowered so that it would expeditiously complete its work in the kernel and return to user mode. This allowed other processes that blocked waiting to enter the kernel to make progress. Once the system call is complete, the process's p_pri is reset to its p_usrpri.
User Mode Priority
The priority that a process within the kernel receives after returning from the blocked queue is called its sleep priority. There is a specific sleep priority associated with every blocking condition. For example, the sleep priority associated with terminal I/O was 28 and disk I/0 was 20.
The user mode scheduling priority depends on three factors:
- the default user mode priority, PUSER, typically 50
- the process's recent CPU usuage, p_cpu
- the nice value assigned by the user, p_nice
The p_usr value is a system-wide default. In most implementations it was 50, indicating the most important level of scheduling for a user process.
Let's be Nice
The p_nice value defaults to 0, but can be increased by users who want to be nice. Remember that the likelihood of a process to be dispatched is inversly proportional to the priority. By increasing the process's nice value, the process is deacreasing its likelihood of being scheduled. Processes are usually "niced" if they are long-running, non-interactive backgorund processes.
Tracking CPU usage
p_cpu is a measure of the process's recent CPU usage. It ranges from 0 - 127 and is initially 0. Ever 1/10th of a second, the ISR that handles clock ticks increments the p_cpu for the current process.
Every 1 second another ISR decreases the p_cpu of all processes, running or not. This reduction is called the decay. SVR3 used a fixed decay of 1/2. The problem with a fixed decay is that it elevates the priority of nearly all processes if the load is very high, since very few processes are getting CPU. This makes the p_cpu field nearly meaningless.
The designers of 4.3BSD remedied this side-effect by using a variable decay that is a fuction of the systems load average, the average number of processes in the run queue over the last second. This formula follows:
decay_factor = (2*load_average)/ (2*load_average + 1)
User Mode Priority - Final Formula
The scheduler computes the process's user priority form these factors as follows:
p_usrpri = PUSER + (p_cpu/4) + (2*p_nice)
Artifacts of the old DEC VAX systems made it much more efficient to collapse the 127 priorities into 32 queues. So these systems in effect had 32 queues each holding processes in a range of 4 priority levels (0-3, 4-7, 8-11, 12-15, 16-19, etc).
The system maintained a 32-bit mask. Each bit represented a single queue. If the bit was set, there were jobs in the queue. If the bit was 0, the queue was empty. The system charged from low-bit to high-bit in this mask until it found a non-empty queue. It would then select a job Round Robin (RR) from this queue to be dispatched.
The round-robin scheduling with a time quantum of 100mS only applied to processes in the same queue. If a process arrived in a lower priority (more important) queue, that process would be scheduled at the end of the currently executing process's quantum.
High priority (less important) processes would not execute until all lower priority (more important) queues were empty.
The queues would be check by means of the bit mask every time a process blocked or a time quantum expired.
This method of scheduling proved viable for general purpose systems, but it does have several limitations:
- it falls apart with a massive number of processes - the overhead of recomputing priorities for every process every second becomes too high
- nice values are a very weak and underutilized way for applications to affect their priorities -- how often doyou nice your jobs?
- A non-preemptive kernel means that important processes may have to wait for lower priority processes in the kernel. The temporary sleep priority is a weak solution.