Static Bus Schedule aware Scratchpad Allocation in Multiprocessors Sudipta Chattopadhyay Abhik Roychoudhury National University of Singapore.

Static Bus Schedule aware Scratchpad Allocation in Multiprocessors

Sudipta Chattopadhyay

Abhik Roychoudhury

National University of Singapore

Scratchpad Memory (Basics)

Scratchpad Memory A fast and software controlled on-chip memory Each memory access is predictable

Problems Cumbersome and error prone if managed by user Need extensive compiler support for automatic management

Scratchpad Allocation

Worst case optimization vs Average case optimization This work is on worst case

ActualBCET

ActualWCET

Execution Time

ObservedWCET

Estimated BCET

Actual

Observed

Over-estimation

WCET = Worst-case Execution TimeBCET = Best-case Execution Time

ObservedBCET

EstimatedWCET

ActualWCET

Scratchpad Allocation

SPM

Task

SPM-0 SPM-n........

Task graph

Task graph

Previous work in our group (Suhendra et. al. RTSS’05)Software cache locking (Puaut ECRTS’07)

Previous work in our group (Suhendra et. al. TOPLAS’10)

SPM-0 SPM-n..……

Task graph

Task graphThis paper with next level of memory accessed by shared bus

An MPSoC Architecture

PE-0 PE-1 PE-N

SPM-0 SPM-1 SPM-N

Shared off-chip data bus

Off-chip memory

External Memory Interface

MPSOC

……

Fast on-chip communication media

SPM architecture

Bypassing memory hierarchy Each memory access is predictable – crucial for time

predictable embedded systems

Non-bypassing memory hierarchy Acts like a fully associative cache Spilling and reloading of memory blocks lead to

unpredictable execution time

Allocation strategy

Consider data memory allocation

Variable locations (private SPM, remote SPM or external memory) are computed at compile time

If two variables share the same space, they are guaranteed to have disjoint lifetime

No reloading cost required

Motivation

Why shared bus makes it different ?

Bus slots for other cores(150 cycles)

m m’

Total delay for ref(m) = (150 + LAT) * freq(m)Total delay for ref(m’) = LAT * freq(m’)

150 * freq(m) > LAT * (freq(m’) – freq(m)) (m likely to reduce WCET more than m’)

this core slot

this core slot

Because shared bus delay is variable

freq(m’) > freq(m)

Motivation Why shared SPM space makes it different ?

SPM - 0 SPM - 1

Not so critical task

An SPM allocator unaware of shared scratchpad space cannot allocate memory blocks accessed in critical tasks to SPM-1

Critical task

Allocator’s View

Motivation Why shared SPM space makes it different ?

SPM - 0 SPM - 1

Critical task

Not so critical task

Allocator’s View

Exploiting shared scratchpad space, more performance can be obtained as the critical tasks can also allocate in SPM-1

Allocation framework

Bus-delay awareWCET analysis

Task WCET

Total delay(bus delay + memory latency)

to access variables along WCEP

WCRT analysis

Variable lifetime and critical path

information

Bus aware SPM allocator

Enough space?

SPM allocation decision

Yes

OptimizedWCRT

No

Application task graph

Allocation framework

Bus-delay awareWCET analysis

Task WCET

Total delay(bus delay + memory latency)

to access variables along WCEP

WCRT analysis

Variable lifetime and critical path

information

Bus aware SPM allocator

Enough space?

SPM allocation decision

Yes

OptimizedWCRT

No

Application task graph

Bus delay aware WCET analysis

Shared bus introduces variable latency for each memory access.

Our previous work approximates the total delay incurred by a static memory reference.

This delay is used as a metric by the greedy SPM allocator.

Allocation framework

Bus-delay awareWCET analysis

Task WCET

Total delay(bus delay + memory latency)

to access variables along WCEP

WCRT analysis

Variable lifetime and critical path

information

Bus aware SPM allocator

Enough space?

SPM allocation decision

Yes

OptimizedWCRT

No

Application task graph

WCRT analysis

t3t2

t4

t1(1)

(2) (2)

(1)

Assigned core

Task graph

Task lifetime : [eStart, lFinish]

eStart(t1) = 0eStart(t4) >= eFinish(t2)eFinish(t4) >= eFinish(t3)

eFinish = eStart + BCET

lStart(t4) >= lFinish (t2)lStart(t4) >= lFinish (t3)

t3 can be preempted by t2

lFinish (t3) = lStart(t3) + WCET(t3) + WCET(t2)+ 2 * BUS_SLOT_LENGTH

Computed WCRT = lFinish(t4)

Earliest timecomputation

Latest timecomputation

All tasks have the same period –the period of the entire task graph

Allocation framework

Bus-delay awareWCET analysis

Task WCET

Total delay(bus delay + memory latency)

to access variables along WCEP

WCRT analysis

Variable lifetime and critical path

information

Bus aware SPM allocator

Enough space?

SPM allocation decision

Yes

OptimizedWCRT

No

Application task graph

Bus aware SPM allocator

Using WCRT analysis we also obtain the lifetime information of each variable

Interference graph Each node is a variable accessed in some task An edge exists between two nodes if their lifetimes interfere

Nodes have weights Higher the total access delay (including bus delay), higher the

weight Higher the weight if accessed in critical path

Bus aware SPM allocatorM1 = 10C1 = 30

C2 = 55

C3 = 10M2 = 10

N = 10

N = 5

C4 = 10M1 = 10

Task T1 in PE-0

C5 = 40

C6 = 55

N = 10

M3 = 10

M3 = 10

Task T2 in PE-1

Assume M1, M2 and M3 are only memory accesses

Critical Task = T1M2 suffers more delay to access than M1.Reduce WCRT by reducing the WCRT of critical task T1

Bus aware SPM allocator

M1 = 10C1 = 30

C2 = 55

C3 = 10M2 = 10

N = 10

N = 5

C4 = 10M1 = 10

Task T1 in PE-0

C5 = 40

C6 = 55

N = 10

M3 = 10

M3 = 10

Task T2 in PE-1

SPM-0

SPM-1

M1 M2

M3

[0,690]

[455, 480]

[375,650]

M2

t = 530

t = 480

T1T2

WCRT = 530 cycles

(empty)

Allocation (iteration 1)

Critical Task = T1

Reduce WCRT by reducing the WCRT of critical task T1

Bus aware SPM allocator

M1 = 10C1 = 30

C2 = 55

C3 = 10M2 = 10

N = 10

N = 5

C4 = 10M1 = 10

Task T1 in PE-0

C5 = 40

C6 = 55

N = 10

M3 = 10

M3 = 10

Task T2 in PE-1

Critical Task = T2

SPM-0

SPM-1

M1 M2

M3

[0,530]

[455, 480]

[375,510]

M2

t = 464

T1 T2

WCRT = 480 cycles

M1 t = 480

Allocation (iteration 2)

Bus aware SPM allocator

M1 = 10C1 = 30

C2 = 55

C3 = 10M2 = 10

N = 10

N = 5

C4 = 10M1 = 10

Task T1 in PE-0

C5 = 40

C6 = 55

N = 10

M3 = 10

M3 = 10

Task T2 in PE-1

M2 and M3 have disjoint lifetimes, allocate same space

SPM-0

SPM-1

M1 M2

M3

[0,464]

[455, 480]

[405,450]

(M2, M3)

t = 464 t = 463

T1 T2

WCRT = 464 cycles

M1

Allocation (iteration 3)

Experimental evaluation

Two real world applications An unmanned aerial vehicle (UAV) controller (papabench) A fragment of an in-orbit spacecraft software (Debie)

Compare WCRT improvement with different SPM size (default: 5% of total data size) Bus slot length (default: 50 cycles to each core) Remote SPM latency (default: 4 cycles)

Compare improvement with bus unaware SPM allocation

WCRT Impr. w.r.t. SPM size

MIS = SPM allocation using our framework

NOBUS = Bus unaware SPM allocator

Improvement over bus unaware allocator = 50%

WCRT impr. w.r.t. bus slot length

Average improvement = 52%

Worst case vs average case

SIM(MIS) - Average case improvement

MIS - Worst case improvement

Allocation framework

Bus-delay awareWCET analysis

Task WCET

Total delay(bus delay + memory latency)

to access variables along WCEP

WCRT analysis

Variable lifetime and critical path

information

Bus aware SPM allocator

Enough space?

SPM allocation decision

Yes

OptimizedWCRT

No

Application task graph

Allocation framework

Bus-delay awareWCET analysis

Task WCET

Total delay(bus delay + memory latency)

to access variables along WCEP

WCRT analysis

Variable lifetime and critical path

information

Bus aware SPM allocator

Enough space?

SPM allocation decision

Yes

OptimizedWCRT

No

Application task graph

Analysisof different

bus arbitration policies

Summary

We have proposed an SPM allocation framework for MPSoCs

Our goal is to reduce the worst case response time (WCRT) of an application

We consider variable bus delays in SPM allocation

Currently, we have the model for TDMA bus only, but the SPM allocation framework can be used for different types of bus arbitration policies