An experimental comparison of different real-time schedulers on multicore systems

Lelli, Juri; Faggioli, Dario; Cucinotta, Tommaso; Lipari, Giuseppe

doi:10.1016/j.jss.2012.05.048

Cited by 30 publications

(23 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…III that is primarily based on message passing, and in which the remaining locks are accessed in pairwise fashion only. 3) We demonstrate that our design performs significantly better at scale than either of the two major current G-EDF implementations by means of average-and worst-case overhead measurements across six multicore platforms with 8,16,24,32, and 64 cores in Sec. IV.…”

Section: B Focus Of This Papermentioning

confidence: 92%

“…IV. In short, we measured the average and maximum runtime overheads of GSN-EDF when managing 8,16,24,32,48, and 64 cores, respectively, of a 2 GHz 64-core Intel Xeon machine. Fig.…”

Section: Litmusmentioning

confidence: 99%

“…The SD patch [26,31,32] for Linux was developed at the Scuola Superiore Sant'Anna in Pisa and fundamentally differs from GSN-EDF in both its design and its goal. While LITMUS RT is primarily a prototyping framework and abstraction layer for experimental real-time research, SD is a comparably mature implementation that directly integrates with the kernel and aims at eventual inclusion into mainline Linux.…”

Section: The Sched Deadline Patch For Linuxmentioning

confidence: 99%

“…We first study two mature, radically different implementations of global real-time scheduling-LITMUS RT 's global earliest-deadline first (G-EDF) scheduler [13,17] and the SCHED DEADLINE scheduler for stock Linux [26,31,32] (SD hereafter)-and show that they indeed do not scale. LITMUS RT 's implementation uses a single coarse-grained lock to protect all scheduler state; its lack of scalability is thus expected.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Scaling global scheduling with message passing

Cerqueira

Vanga

Brandenburg

2014

2014 IEEE 19th Real-Time and Embedded Technology and Applications Symposium (RTAS)

View full text Add to dashboard Cite

Abstract-Global real-time schedulers have earned the reputation of scaling poorly due to the high runtime overheads involved in global state management. In this paper, two mature implementations, one using fine-grained locking (SCHED DEADLINE) and one using coarse-grained locking (LITMUS RT 's G-EDF plugin), are evaluated and it is shown that, regardless of locking granularity, indeed neither scales well w.r.t. worst-case overheads due to excessive lock contention. To demonstrate that this is not an inherent limitation of global scheduling, the design of G-EDF-MP is presented, a global scheduler that uses message passing to avoid lock contention and cache-line sharing. It is shown to offer up to a 23-to 36-fold reduction in worst-case scheduling overhead on a 64-core platform, which translates into much improved schedulability (in some cases, more than 120 additional tasks can be supported).

show abstract

Section: B Focus Of This Papermentioning

confidence: 92%

“…IV. In short, we measured the average and maximum runtime overheads of GSN-EDF when managing 8,16,24,32,48, and 64 cores, respectively, of a 2 GHz 64-core Intel Xeon machine. Fig.…”

Section: Litmusmentioning

confidence: 99%

Section: The Sched Deadline Patch For Linuxmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Scaling global scheduling with message passing

Cerqueira

Vanga

Brandenburg

2014

2014 IEEE 19th Real-Time and Embedded Technology and Applications Symposium (RTAS)

View full text Add to dashboard Cite

show abstract

“…In general, both types of control entities instances are able to instantiate scheduler instances on demand (e.g., in cluster computing environment, there will be one such component per each processor/resource where the application threads are loaded and needed to be executed.). The details of the scheduling mechanism used are presented in [60].…”

Section: Rp or Resource (Information)mentioning

confidence: 99%

ElCore: Dynamic elastic resource management and discovery for future large-scale manycore enabled distributed systems

Zarrin

Aguiar

Barraca

2016

Microprocessors and Microsystems

View full text Add to dashboard Cite

The next generation of many-core enabled large-scale computing systems relies on thousands of billions of heterogeneous processing cores connected to form a single computing unit. In such large-scale computing environments, resource management is one of the most challenging, and complex issues for efficient resource sharing and utilization, particularly as we move toward Future ManyCore Systems (FMCS). This work proposes a novel resource management scheme for future peta-scale many-core-enabled computing systems, based on hybrid adaptive resource discovery, called ElCore. The proposed architecture contains a set of modules which will dynamically be instantiated on the nodes in the distributed system on demand. Our approach provides flexibility to allocate the required set of resources for various types of processes/applications. It can also be considered as a generic solution (with respect to the general requirements of large scale computing environments) which brings a set of interesting features (such as auto-scaling, multitenancy, multi-dimensional mapping, etc,.) to facilitate its easy adaptation to any distributed technology (such as SOA, Grid and HPC many-core). The achieved evaluation results assured the significant scalability and the high-quality resource mapping of the proposed resource discovery and management over highly heterogeneous, hierarchical and dynamic computing environments with respect to several scalability and efficiency aspects while supporting flexible and complex queries with guaranteed discovery results accuracy. The simulation results prove that, using our approach, the mapping between processes and resources can be done with high level of accuracy which potentially leads to a significant enhancement in the overall system performance.

show abstract

Timing analysis of the PREEMPT RT Linux kernel

Oliveira

2015

Softw Pract Exp

View full text Add to dashboard Cite

In the theory of real-time scheduling, tasks are described by mathematical variables, which are used in analytical models in order to prove schedulability of the system. On real-time Linux, tasks are computer programs, and Linux developers try to lower the latencies caused by the Linux kernel, trying to achieve faster response for the highest-priority task. Although both seek temporal correctness, they use different abstractions, which end up separating these efforts in two different worlds, making it hard for the Linux practitioners to understand and apply the formally proved models to the Linux kernel and for theoretical researchers to apply the restrictions imposed by Linux for the theoretical models. This paper traces a parallel between the theory of response-time analysis and the abstractions used in the Linux kernel. The contribution of this paper is threefold. We first identify the PREEMPT RT Linux kernel mechanisms that impact the timing of real-time tasks and map these impacts to the main abstractions used by the real-time scheduling theory. Then, we describe a customized trace tool, based on the existing trace infrastructure of the Linux kernel, that allows the measurement of the delays associated with the main abstractions of the real-time scheduling theory. Finally, we use this customized trace tool to characterize the timing lines resulting from the behavior of the PREEMPT RT Linux kernel.In addition to the function trace, there are other trace plugins, with emphasis on the function graph. The function graph traces the call and return of a function. To improve the understanding of the stack of functions, the output shows the indentation of the functions according to its position in the stack. This is an example of the execution of the function graph:An advantage of the function graph is the ability to determine the execution time of a particular function. It also makes the trace easy to follow, because of the indentation of functions. Ftrace allows the combined use of trace plugins and tracepoints.The tracer proposed in this article, denominated Trace Timeflow, was created based on the function graph tracer, in order to trace the relevant functions. It also uses tracepoints to trace important changes in the system state. Trace timeflowInitially, the new plugin was built as a copy of function graph. From this clone, changes were made to meet our needs. The first change was the fields to be displayed in the trace. Trace format.The format of the trace consists of six fields, as in the following example:The field TASK-PID identifies the task running, it displays the name of the process and its PID. The field PRIO displays the priority. Currently, Linux has 140 priorities, where priority 0 is the highest and 139 is the lowest. The real-time tasks use priorities from 0 to 99, with priorities from 100 to 139 used as time-sharing priorities.The field CPU displays the CPU where the task is running.TIMING ANALYSIS OF THE PREEMPT RT LINUX KERNEL 805 Inside the infinite loop created with for, the function pause(...

show abstract

An experimental comparison of different real-time schedulers on multicore systems

Cited by 30 publications

References 11 publications

Scaling global scheduling with message passing

Scaling global scheduling with message passing

ElCore: Dynamic elastic resource management and discovery for future large-scale manycore enabled distributed systems

Timing analysis of the PREEMPT RT Linux kernel

Contact Info

Product

Resources

About