Reliability model in ns3

classic Classic list List threaded Threaded
4 messages Options
Reply | Threaded
Open this post in threaded view
|

Reliability model in ns3

Vishwesh Rege
Hi,

As part of my research work I was planning to develop a reliability module
in ns3 that tracks the reliability of a node based on temperature, voltage,
frequency and power. My aim is to monitor these parameters, and provide the
values to a mathematical model to estimate the lifetime of a node or
probability of node failure.

Wanted to know if there is any existing implementation or related work on
this. If anyone is currently working on something or has any ideas please
let me know.

Thanks.

- Vishwesh Rege
Reply | Threaded
Open this post in threaded view
|

Re: Reliability model in ns3

Vishwesh Rege
Hi,

Thanks for your feedback.

Currently I plan to have separate models (classes) for temperature, power
and reliability. Voltage and frequency would be fixed attributes at least
initially.

The Temperature model is based on the current temperature and power:
T[k+1] = A T[k] + B P[k+1]

The Power model is based on features such as performance counters,
utilization, etc.
P = sum_i ( a[i] x[i] )

Finally the reliability would be calculated using the values provided by
these models. There is some existing literature on this. But the
implementation might be tricky because of the differing timescales of the
temperature/power models and the reliability model (seconds vs.
days/months).

Also since the EnergySource model already includes voltage attribute for
the node (as you pointed out), implementing dynamic voltage might be
tricky. I'm yet to decide how I would include dynamic models for
voltage/frequency.

If you have any doubts/suggestions do let me know. It would also help me
refine my ideas further.

Thanks,
Vishwesh

On Tue, Apr 18, 2017 at 2:12 PM, Peter Barnes <[hidden email]> wrote:

> This is an interesting idea.  IIRC Vedran was doing some work on failure
> models, which led to adding longer time units, such as years, to Time.
>
> Some thoughts/questions:
>
> • Nodes currently don’t have any of those properties (temperature,
> voltage, frequency, power), out of the box.  How might you go about adding
> them?
>
> •  Power (and voltage, for some EnergySource models) might be available if
> an EnergySource has been aggregated to a Node.  How might you handle a node
> with multiple EnergySources?  (Wait, is that even possible with
> aggregation?  I don’t think so, but it should be…)
>
> °  Along with those attributes we would need dynamical models.  How should
> power dissipation lead to changes in temperature?  Do you also need a model
> (or at least a value) for the ambient temperature?  Do you have to model
> heat sink efficiency in order to get a reasonable temperature rise due to
> the radiated power?  How does over/under-voltage change the power
> dissipation?
>
> Peter
>
> On Apr 18, 2017, at 11:29 AM, Vishwesh Rege <[hidden email]> wrote:
>
> Hi,
>
> As part of my research work I was planning to develop a reliability module
> in ns3 that tracks the reliability of a node based on temperature, voltage,
> frequency and power. My aim is to monitor these parameters, and provide the
> values to a mathematical model to estimate the lifetime of a node or
> probability of node failure.
>
> Wanted to know if there is any existing implementation or related work on
> this. If anyone is currently working on something or has any ideas please
> let me know.
>
> Thanks.
>
> - Vishwesh Rege
>
>
> ____________
> Peter Barnes
> [hidden email]
>
>
Reply | Threaded
Open this post in threaded view
|

Re: Reliability model in ns3

Vedran Miletić-2
Hi Vishwesh,

what's the state of this work?

Regards,
Vedran

On 04/21/2017 08:27 AM, Vishwesh Rege wrote:

> Hi,
>
> Thanks for your feedback.
>
> Currently I plan to have separate models (classes) for temperature, power
> and reliability. Voltage and frequency would be fixed attributes at least
> initially.
>
> The Temperature model is based on the current temperature and power:
> T[k+1] = A T[k] + B P[k+1]
>
> The Power model is based on features such as performance counters,
> utilization, etc.
> P = sum_i ( a[i] x[i] )
>
> Finally the reliability would be calculated using the values provided by
> these models. There is some existing literature on this. But the
> implementation might be tricky because of the differing timescales of the
> temperature/power models and the reliability model (seconds vs.
> days/months).
>
> Also since the EnergySource model already includes voltage attribute for
> the node (as you pointed out), implementing dynamic voltage might be
> tricky. I'm yet to decide how I would include dynamic models for
> voltage/frequency.
>
> If you have any doubts/suggestions do let me know. It would also help me
> refine my ideas further.
>
> Thanks,
> Vishwesh
>
> On Tue, Apr 18, 2017 at 2:12 PM, Peter Barnes <[hidden email]> wrote:
>
>> This is an interesting idea.  IIRC Vedran was doing some work on failure
>> models, which led to adding longer time units, such as years, to Time.
>>
>> Some thoughts/questions:
>>
>> • Nodes currently don’t have any of those properties (temperature,
>> voltage, frequency, power), out of the box.  How might you go about adding
>> them?
>>
>> •  Power (and voltage, for some EnergySource models) might be available if
>> an EnergySource has been aggregated to a Node.  How might you handle a node
>> with multiple EnergySources?  (Wait, is that even possible with
>> aggregation?  I don’t think so, but it should be…)
>>
>> °  Along with those attributes we would need dynamical models.  How should
>> power dissipation lead to changes in temperature?  Do you also need a model
>> (or at least a value) for the ambient temperature?  Do you have to model
>> heat sink efficiency in order to get a reasonable temperature rise due to
>> the radiated power?  How does over/under-voltage change the power
>> dissipation?
>>
>> Peter
>>
>> On Apr 18, 2017, at 11:29 AM, Vishwesh Rege <[hidden email]> wrote:
>>
>> Hi,
>>
>> As part of my research work I was planning to develop a reliability module
>> in ns3 that tracks the reliability of a node based on temperature, voltage,
>> frequency and power. My aim is to monitor these parameters, and provide the
>> values to a mathematical model to estimate the lifetime of a node or
>> probability of node failure.
>>
>> Wanted to know if there is any existing implementation or related work on
>> this. If anyone is currently working on something or has any ideas please
>> let me know.
>>
>> Thanks.
>>
>> - Vishwesh Rege
>>
>>
>> ____________
>> Peter Barnes
>> [hidden email]
>>
>>

--
Vedran Miletić
vedran.miletic.net
Reply | Threaded
Open this post in threaded view
|

Re: Reliability model in ns3

Vishwesh Rege
Hi,

I have developed models for temperature and power which are quite
straightforward. The power model is based only on the current frequency,
and the temperature model is based on the previous temperature and power
consumption. The reliability model work is still in progress and I'm yet to
validate all these models. Also the models are assuming CPU frequency,
voltage, etc. and not currently taking into account the radio.
High level call sequence:
Each Time unit --> PowerModel --> TemperatureModel --> ReliabilityModel

You can contact me if you have any other questions.

Thanks,
Vishwesh


On Thu, Jul 27, 2017 at 6:31 AM, Vedran Miletić <[hidden email]> wrote:

> Hi Vishwesh,
>
> what's the state of this work?
>
> Regards,
> Vedran
>
> On 04/21/2017 08:27 AM, Vishwesh Rege wrote:
> > Hi,
> >
> > Thanks for your feedback.
> >
> > Currently I plan to have separate models (classes) for temperature, power
> > and reliability. Voltage and frequency would be fixed attributes at least
> > initially.
> >
> > The Temperature model is based on the current temperature and power:
> > T[k+1] = A T[k] + B P[k+1]
> >
> > The Power model is based on features such as performance counters,
> > utilization, etc.
> > P = sum_i ( a[i] x[i] )
> >
> > Finally the reliability would be calculated using the values provided by
> > these models. There is some existing literature on this. But the
> > implementation might be tricky because of the differing timescales of the
> > temperature/power models and the reliability model (seconds vs.
> > days/months).
> >
> > Also since the EnergySource model already includes voltage attribute for
> > the node (as you pointed out), implementing dynamic voltage might be
> > tricky. I'm yet to decide how I would include dynamic models for
> > voltage/frequency.
> >
> > If you have any doubts/suggestions do let me know. It would also help me
> > refine my ideas further.
> >
> > Thanks,
> > Vishwesh
> >
> > On Tue, Apr 18, 2017 at 2:12 PM, Peter Barnes <[hidden email]> wrote:
> >
> >> This is an interesting idea.  IIRC Vedran was doing some work on failure
> >> models, which led to adding longer time units, such as years, to Time.
> >>
> >> Some thoughts/questions:
> >>
> >> • Nodes currently don’t have any of those properties (temperature,
> >> voltage, frequency, power), out of the box.  How might you go about
> adding
> >> them?
> >>
> >> •  Power (and voltage, for some EnergySource models) might be available
> if
> >> an EnergySource has been aggregated to a Node.  How might you handle a
> node
> >> with multiple EnergySources?  (Wait, is that even possible with
> >> aggregation?  I don’t think so, but it should be…)
> >>
> >> °  Along with those attributes we would need dynamical models.  How
> should
> >> power dissipation lead to changes in temperature?  Do you also need a
> model
> >> (or at least a value) for the ambient temperature?  Do you have to model
> >> heat sink efficiency in order to get a reasonable temperature rise due
> to
> >> the radiated power?  How does over/under-voltage change the power
> >> dissipation?
> >>
> >> Peter
> >>
> >> On Apr 18, 2017, at 11:29 AM, Vishwesh Rege <[hidden email]>
> wrote:
> >>
> >> Hi,
> >>
> >> As part of my research work I was planning to develop a reliability
> module
> >> in ns3 that tracks the reliability of a node based on temperature,
> voltage,
> >> frequency and power. My aim is to monitor these parameters, and provide
> the
> >> values to a mathematical model to estimate the lifetime of a node or
> >> probability of node failure.
> >>
> >> Wanted to know if there is any existing implementation or related work
> on
> >> this. If anyone is currently working on something or has any ideas
> please
> >> let me know.
> >>
> >> Thanks.
> >>
> >> - Vishwesh Rege
> >>
> >>
> >> ____________
> >> Peter Barnes
> >> [hidden email]
> >>
> >>
>
> --
> Vedran Miletić
> vedran.miletic.net
>