A task remapping technique for reliable multi-core embedded systems

Abstract

With the continuous scaling of semiconductor technology, the life-time of circuit is decreasing so that processor failure becomes an important issue in MPSoC design. A software solution to tolerate run-time processor failure is to migrate tasks from the failed processors to the live processors when failure occurs. Previous works on run-time task migration usually aim to minimize the migration overhead with or without a given latency constraint. For streaming applications, however, it is more important to minimize the throughput degradation than the migration overhead or the latency. Hence, we propose a task remapping technique to minimize the throughput degradation assuming that the migration overhead can be amortized safely. The target multi-core system assumed in this paper consists of processor pools and each pool consists of homogeneous processors. The proposed technique is based on an intensive compile-time analysis for all possible failure scenarios. It involves the following steps; 1) Determine the static mapping of tasks onto the live processors, aiming to minimize the throughput degradation: 2) Find an optimal processor-to-processor mapping to minimize the task migration overhead: and 3) Store the resultant task remapping information that includes task mapping and processor-to-processor mapping results. Since the task remapping information is pre-computed at compile-time for all possible failure scenarios, it should be efficiently represented and stored. At run-time, we simply remap the tasks following the compile-time decision. We examine the scalability of the proposed technique on both space and run-time overhead for compile-time analysis varying the number of failed processors. Through intensive experiments, we show that the proposed technique outperforms the previous works with respect to application throughput.

Control-Quality Optimization for Distributed Embedded Systems with Adaptive Fault Tolerance

“…This information is stored in the nodes of the platform [11], [2]. Another phase during system reconfiguration is task migration [8] that takes place when tasks running on failed nodes must be activated at other nodes in the system. The system has the ability to migrate tasks to other nodes in the platform [8].…”

“…Another phase during system reconfiguration is task migration [8] that takes place when tasks running on failed nodes must be activated at other nodes in the system. The system has the ability to migrate tasks to other nodes in the platform [8]. Each node stores information regarding those tasks that it must migrate on the bus when the system is adapting to a new configuration.…”

“…It is thus important to construct systems that are resilient to node failures and, in addition, provide as high control quality as possible with the available computation and communication resources during operation. Related approaches for embedded systems design and permanent faults, relying on task remapping by replication and migration, have been proposed recently [7], [8]. These methods are restricted to a predefined set of fault scenarios to be handled at runtime with no requirements on system performance.…”

Analysis and optimization of fault-tolerant task scheduling on multiprocessor embedded systems

“…The ideal case is that the system recovers from faults with only minor re-configuration. Since attaining the optimal task migration decision is a highly complex task, recent work [13] proposes to compute the task re-mappings statically offline and store them in tables. The pre-computed configurations are then applied at runtime if a processor fails.…”

“…Naturally, task migration is only possible if adequate hardware resources are still available. Re-mapping schemes are usually designed carefully to guarantee the feasibility and minimize the migration cost [13,25].…”