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

▶
the real time multi resource task model
SMART_READER_LITE
LIVE PREVIEW

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

Dec 12, 2023 221 likes •457 views

The Real-Time Multi-Resource Task Model The Real-Time Multi-Resource Task Model RTSOPS12 Pisa, Italy July 10 th , 2012 Cong Liu The University of North Carolina at Chapel Hill UNC Chapel Hill

slide-1
SLIDE 1

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

The Real-Time Multi-Resource Task Model

Cong Liu The University of North Carolina at Chapel Hill

RTSOPS’12 Pisa, Italy July 10th, 2012

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

slide-2
SLIDE 2

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Heterogeneous Systems for Real-Time Computing

2

  • Integrate additional specialized resources such as FPGA, GPU

➡

high performance

➡

energy efficiency

➡

highly reactive systems that interact with other environments

slide-3
SLIDE 3

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Heterogeneous Systems for Real-Time Computing

2

  • 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

slide-4
SLIDE 4

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Real-Time Sporadic Multi-Resource (SMR) Task Model

3

  • 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

slide-5
SLIDE 5

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Real-Time Sporadic Multi-Resource (SMR) Task Model

4

CPU GPU FPGA CPU

τ1

slide-6
SLIDE 6

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Real-Time Sporadic Multi-Resource (SMR) Task Model

5

τ11 (e11=1)

CPU

τ1

slide-7
SLIDE 7

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Real-Time Sporadic Multi-Resource (SMR) Task Model

6

τ11 (e11=1)

CPU GPU

τ12 (e12=2)

τ1

slide-8
SLIDE 8

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Real-Time Sporadic Multi-Resource (SMR) Task Model

7

τ11 (e11=1)

CPU GPU FPGA

τ12 (e12=2) τ13 (e13=1)

τ1

slide-9
SLIDE 9

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Real-Time Sporadic Multi-Resource (SMR) Task Model

8

τ11 (e11=1)

CPU GPU FPGA CPU

τ14 (e14=1) τ12 (e12=2) τ13 (e13=1)

τ1

slide-10
SLIDE 10

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Real-Time Sporadic Multi-Resource (SMR) Task Model

9

τ11 (e11=1)

CPU GPU FPGA CPU

τ14 (e14=1) τ12 (e12=2) τ13 (e13=1)

p1 = d1 = 6

τ1

slide-11
SLIDE 11

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

Real-Time Sporadic Multi-Resource (SMR) Task Model

10

τ11 (e11=1)

CPU GPU FPGA CPU

τ14 (e14=1) τ12 (e12=2) τ13 (e13=1)

p1 = d1 = 6

τ1

➡ Utilization on CPU = 2 / 6

slide-12
SLIDE 12

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Scheduling Restrictions on Certain Resources

Sensors Sensors

11

  • Non-preemptivity or non-job-migration

restrictions

➡ GPU: non-preemptive ➡ Job migrations may cause significant overheads on

resources such as GPU and FPGA

slide-13
SLIDE 13

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

The Open Problem

Sensors Sensors

12

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?

slide-14
SLIDE 14

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

13

The Challenge

  • Difficult due to precedence constraints and interferences

between executions of tasks on multiple resources

slide-15
SLIDE 15

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

14 1 2 3 4 5 6 7 8 9 10

4 1 1 1 5 1

job release job deadline

R1

τ1 τ2

R2 R1 R1 R2 R1

The Challenge

➡ Two SMR tasks executed on two resources, where R1 has two

processors and R2 has one processor

➡ Lightly loaded system: 0.8 and 0.5 for the total utilization on R1

and R2, respectively

slide-16
SLIDE 16

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

15 1 2 3 4 5 6 7 8 9 10

4 1 1 1 5 1

job release job deadline

R1

τ1 τ2

R2 R1 R1 R2 R1

The Challenge

➡ Two SMR tasks executed on two resources, where R1 has two

processors and R2 has one processor

➡ Lightly loaded system: 0.8 and 0.5 for the total utilization on R1

and R2, respectively

slide-17
SLIDE 17

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

15 1 2 3 4 5 6 7 8 9 10

4 1 1 1 5 1

job release job deadline

R1

τ1 τ2

R2 R1 R1 R2 R1

The Challenge

➡ Two SMR tasks executed on two resources, where R1 has two

processors and R2 has one processor

➡ Lightly loaded system: 0.8 and 0.5 for the total utilization on R1

and R2, respectively

slide-18
SLIDE 18

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Sensors Sensors

16

The precedence constraints among phases belonging to the same SMR task plus the interferences brought by

  • ther tasks quite negatively impact schedulability

The Challenge

1 2 3 4 5 6 7 8 9 10

4 1 1 1 5 1

job release job deadline

R1

τ1 τ2

R2 R1 R1 R2 R1

slide-19
SLIDE 19

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

An Intuitive Approach: Assigning Intermediate Releases and Deadlines

Sensors Sensors

17

  • 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

slide-20
SLIDE 20

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

An Example

Sensors Sensors

18 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 R1 R2 R1 R2 R3

2 2 2 4 2

τ1 τ2

job release job deadline

slide-21
SLIDE 21

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

An Example

Sensors Sensors

19 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 R1 R2 R1 R2 R3

2 2 2 4 2

τ1 τ2

Phase release Phase deadline

➡ If all phases of an SMR task can meet their assigned

intermediate deadlines, they become independent from each other

slide-22
SLIDE 22

UNC Chapel Hill C. Liu The Real-Time Multi-Resource Task Model

Other Insights

Sensors Sensors

20

  • Study special cases

➡ SMR task systems that only request two resources each

  • f which contains a single processor

➡ Execution times of all phases of all tasks are identical