Real-Time Systems Group Accounting for Cache Related Pre-emption Delays in Hierarchal Scheduling Will Lunniss 1 Sebastian Altmeyer 2 Giuseppe Lipari 3,4.

Real-Time Systems Group

Accounting for Cache Related Pre-emption Delays in Hierarchal Scheduling

Will Lunniss1 Sebastian Altmeyer2 Giuseppe Lipari3,4 Robert I. Davis1

1Department of Computer Science, University of York, UK{wl510, rob.davis}@york.ac.uk

2Computer Systems Architecture Group, University of Amsterdam, [email protected]

3Scuola Superiore, Sant'Anna, [email protected]

4LSV – ENS Cachan, [email protected]

http://www.york.ac.uk/



2

Outline

Hierarchical Scheduling Cache Related Pre-emption Delays (CRPD) Accounting for CRPD in Hierarchical Scheduling Evaluation Conclusions




3

Outline





4

Hierarchical Scheduling

Hierarchical scheduling provides a means of composing multiple real-time components (applications) on a single processor

Components should be isolated and not interfere with each other

Divide up access to the processor using a global scheduler

Schedule tasks in each component using a local scheduler




5


Assume single processor Assume each component has a dedicated server, S,

that allocates processor capacity to it Assume servers are scheduled globally using non-pre-

emptive scheduler Assume tasks are scheduled locally using pre-emptive

FP scheduler Each server SG, has a budget QG, and a period PG




6


Components scheduled globally using non pre-emptive scheduling

Tasks scheduled locally using pre-emptive FP scheduling




7

Supply Bound Function

Accounts for non dedicated access to the processor inverse supply bound function (isbf)

determine the maximum amount of time needed by a specific server to supply some capacity c

[1] I. Shin and I. Lee, "Periodic Resource Model for Compositional Real-Time Guarantees," in Proceedings of the 24th IEEE Real-Time Systems Symposium (RTSS), Cancun, Mexico, 2003, pp. 2-13.




8

Schedulability Analysis for Hierarchical Systems

j

iGhpj j

jii

Gi C

T

JRCisbfR

),(

1

Wrap response time analysis in the inverse supply bound function

Restrict the set of tasks to those in the same component




9

Outline





10

Cache Related Pre-emption Delays (CRPD)

Pre-empting task can evict cache blocks belonging to the pre-empted task

CRPD are introduced when the pre-empted task has to reload some of those evicted cache blocks after resuming

Tasks in other components could evict cache blocks, causing ‘inter-component’ CRPD




11

Cache Block Categorisation

Evicting Cache Blocks (ECBs) Loaded into cache and can therefore evict other blocks

Useful Cache Blocks (UCBs) Reused once they have been loaded into cache before potentially

being evicted by the task

If evicted by another task, they may have to be reloaded which introduces CRPD

UCBs are always also ECBs




12

CRPD in Hierarchical Scheduling

Component-level CRPD

Task-level CRPD




13

Outline





14

Integrating CRPD Analysis

Gi

iGhpjjij

j

jii

Gi C

T

JRCisbfR

),(,

1

Task-level CRPD determined using the Combined Multiset approach [1]

Component-level CRPD

[1] Altmeyer, S., Davis, R.I., and Maiza, C. Improved Cache Related Pre-emption Delay Aware Response Time Analysis for Fixed Priority Pre-emptive Systems. Real-Time Systems, 48, 5 (September 2012), 499-512.




15

Approach Variation 1 ‘-All’

Assume that every time server SG is suspended, all of the other servers run and their tasks evict all the cache blocks that they use

This approach could be pessimistic if servers do not have equal periods




16

Approach Variation 1 ‘-All’

Assume that every time server SG is suspended, all of the other servers run and their tasks evict all the cache blocks that they use

G

ii

G

P

RRE 1




17

UCB-ECB-ALL

GZ

ZiGk

kiGG

i RE Z

,hep

ECBUCB BRT

Calculate the union of ECBs of all other components

Calculate the union of UCBs of task τi and all tasks that are part of component G that could pre-empt it

Multiplied by the number of times component G is suspended and resumed during the response time of task τi

Multiplied by the cost incurred when reloading a block, the Block Reload Time (BRT)




18

Approach Variation 2 ‘-Counted’

Consider each component separately by calculating the number of disruptive executions that server SZ can have on task τi in component G during the response time of task τi

Tasks in component Z cannot evict the same cache block multiple times even if component Z runs more than once while component G is suspended




19

Approach Variation 2 ‘-Counted’

Consider each component separately by calculating the number of disruptive executions that server SZ can have on task τi in component G during the response time of task τi

ZGGZ

P

ttEtSX 1,min,




20

UCB-ECB-Counted

Calculate the union of UCBs of task τi and all tasks that are part of component G that could pre-empt it, intersected with the ECBs of component Z

Multiplied by the number of disruptive executions of component Z during the response time of task τi

GZZ iGk

kiGZ

Gi

RSX Z

,hep

ECBUCB,

BRT

Summed for all other components




21

Approach Variation 3 ‘-Open’

Ignore details of other components, assume all blocks are evicted by the other components

Variation of ‘-All’ approaches

The union of ECBs of the other components is assumed to contain all cache blocks

Allows the analysis to be used when no information about the other components is available




22

Outline





23

Evaluation Methodology

Assign equal server periods to all components Split tasks equally between components

Default of 24 tasks and 3 components

Determine if the system is schedulable Perform a binary search to find the minimum server capacity required

to schedule each component

As server periods are equal, if the total capacity of all servers is less than 100% then the system is schedulable




24

Evaluation

Considering just the other components, and doing so individually performs the worst

Accounting for all of the other components at once, or assuming that the cache is flushed, performs the best




25

Number of Components

Assuming the cache if flushed every time a component is suspended is not effective with one component

There is no difference between the ‘-All’ and ‘-Counted’ variants when there are two components




26

Server PeriodMinimising the server periods maximises schedulability when component CRPD is not considered

In practice, the server period must be carefully selected




27

Cache Utilisation

When the cache utilisation is low, not considering the other components can be overly pessimistic

Once the cache utilisation of the other components is > 1, assuming the cache is flushed each time a component is suspended can be as effective as the more sophisticated approach




28

Outline





29

Conclusions

Presented a number of approaches for bounding CRPD under hierarchical scheduling with a local FP scheduler

Shown that it is important to consider inter-component CRPD

Note that minimising server periods does not maximise schedulability, as is the case when inter-component CRPD is not considered

Must select the server period carefully

If the components fill the cache, simply assuming the cache is flushed each time a component is suspended can be very effective with equal server periods

Allows a bound on CRPD to be calculated without any information about the other components in the system




30

Thank you for listening

Any Questions?



Real-Time Systems Group Accounting for Cache Related Pre-emption Delays in Hierarchal Scheduling Will Lunniss 1 Sebastian Altmeyer 2 Giuseppe Lipari 3,4.

Documents

hierarchical systems

evicted cache blocks

hierarchal scheduling

crpd ucbs

preempted task crpd

pre emptive scheduling

intercomponent crpd

single processor components