Multi-Core Software

Process Scheduling Challenges in the Era of Multi-Core Processors

MULTI-CORE AWARE LINUX* PROCESS SCHEDULER

In this section, we consider the Linux* operating system as an example and see how some of these scheduling challenges are addressed. A new scheduler domain representing multi-core characteristics has been added to the domain hierarchy of the Linux process scheduler. This scheduler domain helps identify cores sharing the same package and sharing resources, and it paves the way for the multi-core scheduler enhancements.

Figure 5: Scheduling mechanism showing four running tasks scheduled on four L2s on a dual package with Intel Core 2 Quad processors
click image for larger view
 

By default, the current Linux kernel scheduler distributes the running tasks equally among the available last-level caches in an SMP domain. Within logical CPUs that share the last-level cache, the scheduler distributes the load equally, first among the available CPU cores and then among the available logical thread siblings. For example, consider a dual package SMP platform with Intel® Core™2 quad processors with four running tasks. The multi-core-aware Linux process scheduler distributes these four running tasks among the four L2's that are available in the system as shown in Figure 5. This scheduling mechanism will lead to maximized resource utilization and minimized resource contention. And as observed in the previous sections, this will lead to optimal performance for most of the workloads. On platforms with dynamic acceleration technology, this mechanism will also result in optimal performance by making the cores run faster.

Figure 6: Scheduling mechanism showing four running tasks scheduled on four cores in one single package on a dual package with Intel Core 2 Quad processors
click image for larger view
 

For optimal power savings or for workloads that benefit most by running the tasks on the cores sharing resources, the Linux kernel provides a tunable that can be set by an administrator. When this tunable is set, the process scheduler will try to minimize the packages in an SMP domain that carry the load. It will first try to load all the logical threads and cores in the package before distributing the load to another package. This policy is referred to as a power-savings policy. For example, consider the same four-task scenario on a dual package SMP platform with Intel Core 2 Quad processors. With the power-savings tunable set, all the four tasks will be run on the four cores residing in a single package as shown in Figure 6. Minimizing the number of packages that are active will lead to optimal power savings. As seen before, in the absence of dynamic acceleration support, this scheduling mechanism will not have any impact on performance for workloads such as the IMDS workload considered in Table 1.

Scheduling policy decisions are left to the administrator in the hope that the target workloads will be analyzed offline and the tunable will be set based on optimal performance and/or power-savings requirements. By default, the process scheduler takes a non-aggressive approach when distributing the load among the available shared resources.

Back to Top