User manual NOVELL LINUX ENTERPRISE REAL TIME 10 SP1 QUICK START

[. . . ] SUSE Linux Enterprise Real Time 10 SP1 Quick Start SUSE Linux Enterprise 10 SP1 NOVELL® QUICK START CARD SUSE Linux Enterprise Real Time is an add-on to SUSE® Linux Enterprise that allows you to run tasks which require deterministic real-time processing, in a SUSE Linux Enterprise environment. SUSE Linux Enterprise Real Time meets this requirement by offering several different options for CPU and IO scheduling, CPU shielding and setting CPU affinities to processes. Dedicated CPUs, together with some predefined memory, work on a number of tasks. To retrieve the cpuset of a specific task with a certain process id pid, use the command cat /proc/pid/cpuset. [. . . ] If a task has lots of CPU power available, but must wait for the disk, it will not work as efficient as it could. The Linux kernel provides three different scheduling classes to determine the I/O handling for a process. The Best Effort Scheduler The Best Effort scheduler is the default I/O scheduler, and is used for all processes that do not specify a different I/O scheduler class. By default, this scheduler sets its niceness according to the nice value of the running process. The lowest priority is represented by a nice level of seven, the highest priority is zero. The Real Time Scheduler The real-time I/O class always gets the highest priority for disk access. The other schedulers will only be served, if no real-time request is present. This scheduling class may easily lock up the system if not implemented with care. The real-time scheduler defines nice levels just like the Best Effort scheduler. The Idle Scheduler The Idle scheduler does not define any nice levels. I/O is only done in this class, if no other scheduler runs an I/O request. This scheduler has the lowest available priority and can be used for processes that are not timecritical at all. This provides a means to tune the scheduler of 3 Specifying a CPU Affinity with taskset The default behavior of the kernel, is to keep a process running on the same CPU, if the system load is balanced over the available CPUs. Otherwise, the kernel tries to improve the load balancing by moving processes to an idling CPU. In some situations, however, it is desirable to set a CPU affinity for a given process. In this case, the kernel will not move the process away from the selected CPUs. For example, if you use shielding, the shielded CPUs will not run any process that does not have an affinity to the shielded CPUs. Another possibility is to run all low priority tasks on a selected CPU to remove load from the other CPUs. Note, that if a task is running inside a specific cpuset, the affinity mask must match at least one of the CPUs available in this set. The taskset command will not move a process outside the cpuset it is running in. To set or retrieve the CPU affinity of a task, a bitmask is used, that is represented by a hexadecimal number. If you count the bits of this bitmask, the lowest bit represents the first logical CPU as they are found in /proc/cpuinfo. [. . . ] write_starved In case some read or write requests hit their deadline, the scheduler prefers the read requests by default. To prevent write requests from being postponed forever, the variable write_starved controls how often read requests are preferred until write requests are preferred over read requests. CFQ Scheduler back_seek_max and back_seek_penalty The CFQ scheduler normally uses a strict ascending elevator. When needed, it also allows small backward seeks, but it puts some penalty on them. [. . . ]