The Real-Time Multi-Resource Task Model RTSOPS12 Pisa, Italy July - PowerPoint PPT Presentation

The Real-Time Multi-Resource Task Model The Real-Time Multi-Resource Task Model RTSOPS’12 Pisa, Italy July 10 th , 2012 Cong Liu The University of North Carolina at Chapel Hill UNC Chapel Hill C. Liu Work supported by NSF grants CNS 0834270, CNS 0834132, and CNS 1016954; ARO grant W911NF-09-1-0535; AFOSR grant FA9550-09-1-0549; and AFRL grant FA8750-11-1-0033

The Real-Time Multi-Resource Task Model Heterogeneous Systems for Real-Time Computing • Integrate additional specialized resources such as FPGA, GPU high performance ➡ energy efficiency ➡ highly reactive systems that interact with other environments ➡ 2 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Heterogeneous Systems for Real-Time Computing • Integrate additional specialized resources such as FPGA, GPU high performance ➡ energy efficiency ➡ highly reactive systems that interact with other environments ➡ • IBM cell architecture One “general” processor with ➡ eight “specialized” processors Specialized processors designed ➡ for handling vectorized floating point code execution 2 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Real-Time Sporadic Multi-Resource (SMR) Task Model • Model real-time tasks that access multiple resources during execution ➡ Extend the sporadic task model ➡ Sporadically release jobs ➡ Each job contains several phases, each executed on a specific resource 3 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Real-Time Sporadic Multi-Resource (SMR) Task Model CPU GPU FPGA CPU τ 1 4 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Real-Time Sporadic Multi-Resource (SMR) Task Model CPU τ 1 τ 11 (e 11 =1) 5 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Real-Time Sporadic Multi-Resource (SMR) Task Model CPU GPU τ 1 τ 11 (e 11 =1) τ 12 (e 12 =2) 6 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Real-Time Sporadic Multi-Resource (SMR) Task Model CPU GPU FPGA τ 1 τ 11 (e 11 =1) τ 12 (e 12 =2) τ 13 (e 13 =1) 7 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Real-Time Sporadic Multi-Resource (SMR) Task Model CPU GPU FPGA CPU τ 1 τ 11 (e 11 =1) τ 12 (e 12 =2) τ 13 (e 13 =1) τ 14 (e 14 =1) 8 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Real-Time Sporadic Multi-Resource (SMR) Task Model CPU GPU FPGA CPU τ 1 τ 11 (e 11 =1) τ 12 (e 12 =2) τ 13 (e 13 =1) τ 14 (e 14 =1) p 1 = d 1 = 6 9 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Real-Time Sporadic Multi-Resource (SMR) Task Model CPU GPU FPGA CPU τ 1 τ 11 (e 11 =1) τ 12 (e 12 =2) τ 13 (e 13 =1) τ 14 (e 14 =1) p 1 = d 1 = 6 ➡ Utilization on CPU = 2 / 6 10 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Scheduling Restrictions on Certain Resources • Non-preemptivity or non-job-migration restrictions ➡ GPU: non-preemptive ➡ Job migrations may cause significant overheads on resources such as GPU and FPGA 11 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model The Open Problem How to schedule a set of hard real-time SMR tasks on a heterogeneous platform consisting of multiple types of resources, where each type of resource may have multiple processors and require certain scheduling restrictions? 12 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model The Challenge • Difficult due to precedence constraints and interferences between executions of tasks on multiple resources 13 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model The Challenge job release job deadline R 1 R 2 R 1 τ 1 1 4 1 R 1 R 2 R 1 τ 2 1 1 5 0 1 2 3 4 5 6 7 8 9 10 ➡ Two SMR tasks executed on two resources, where R 1 has two processors and R 2 has one processor ➡ Lightly loaded system: 0.8 and 0.5 for the total utilization on R 1 and R 2 , respectively 14 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model The Challenge job release job deadline R 1 R 2 R 1 τ 1 1 4 1 R 1 R 2 R 1 τ 2 1 1 5 0 1 2 3 4 5 6 7 8 9 10 ➡ Two SMR tasks executed on two resources, where R 1 has two processors and R 2 has one processor ➡ Lightly loaded system: 0.8 and 0.5 for the total utilization on R 1 and R 2 , respectively 15 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model The Challenge job release job deadline R 1 R 2 R 1 τ 1 1 4 1 R 1 R 2 R 1 τ 2 1 1 5 0 1 2 3 4 5 6 7 8 9 10 ➡ Two SMR tasks executed on two resources, where R 1 has two processors and R 2 has one processor ➡ Lightly loaded system: 0.8 and 0.5 for the total utilization on R 1 and R 2 , respectively 15 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model The Challenge job release job deadline R 1 R 2 R 1 τ 1 1 4 1 R 1 R 2 R 1 τ 2 1 1 5 0 1 2 3 4 5 6 7 8 9 10 The precedence constraints among phases belonging to the same SMR task plus the interferences brought by other tasks quite negatively impact schedulability 16 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model An Intuitive Approach: Assigning Intermediate Releases and Deadlines • Each phase of an SMR task is transformed into a constrained-deadline subtask ➡ Each subtask requests a single resource ➡ Apply corresponding existing schedulability tests on each resource ➡ The original SMR task system is schedulable if the transformed subtasks on every resource are schedulable 17 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model An Example job release job deadline R 1 R 2 τ 1 2 2 R 1 R 2 R 3 τ 2 2 4 2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 18 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model An Example Phase release Phase deadline R 1 R 2 τ 1 2 2 R 1 R 2 R 3 τ 2 2 4 2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ➡ If all phases of an SMR task can meet their assigned intermediate deadlines, they become independent from each other 19 UNC Chapel Hill C. Liu Sensors Sensors

The Real-Time Multi-Resource Task Model Other Insights • Study special cases ➡ SMR task systems that only request two resources each of which contains a single processor ➡ Execution times of all phases of all tasks are identical 20 UNC Chapel Hill C. Liu Sensors Sensors