Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.arch.embedded > #24903 > unrolled thread

Instrumentation question.

Back to article view | Back to comp.arch.embedded

Contents

  Instrumentation question. Les Cargill <lcargill99@comcast.com> - 2016-09-07 00:50 -0500
    Re: Instrumentation question. Paul Rubin <no.email@nospam.invalid> - 2016-09-07 00:05 -0700
      Re: Instrumentation question. Les Cargill <lcargill99@comcast.com> - 2016-09-07 13:07 -0500
    Re: Instrumentation question. Tom Gardner <spamjunk@blueyonder.co.uk> - 2016-09-07 08:49 +0100
      Re: Instrumentation question. Les Cargill <lcargill99@comcast.com> - 2016-09-07 13:10 -0500
        Re: Instrumentation question. Tom Gardner <spamjunk@blueyonder.co.uk> - 2016-09-07 23:20 +0100

#24903 — Instrumentation question.

I've always suspected there was something clever that I could use for a 
thing like this, but ... I was too afraid to ask about it.

Suppose I am on an extremely deterministic platform.

Suppose I can calculate the number of microsconds it takes to call 
something[1].

Suppose I can check and calculate things about this count every <n> 
seconds.

Suppose also that there are two cases for "something" :

- There is nothing to do.
- There is something to do. "Something" is constrained to "one thing" by 
design.

Let's also assume I can (cheaply) calculate the minimum and maximum 
number of microseconds to do either of "nothing" or "something".

Obviously a histogram is in order. But wait!.

It seems to me that I should be able to estimate a histogram based on:
- minimum time
- maximum time
- average time.

What is this sort of ... inquiry called, and how do I stop being quite
so dull about it?

[1] I suspect that one answer is to extend the number of buckets
into more finely grained measurements of sub-operations of "something".
Example: a check for "serial port has data" vs. "relieve one unit from 
serial port" vs the overhead from all the dance surrounding those two.

--
Les Cargill

[toc] | [next] | [standalone]

#24904

Les Cargill <lcargill99@comcast.com> writes:
> Obviously a histogram is in order. But wait!.
> It seems to me that I should be able to estimate a histogram based on:
> - minimum time
> - maximum time
> - average time.
> What is this sort of ... inquiry called, and how do I stop being quite
> so dull about it?

You want a probability distribution of the time it takes to do an
operation?

You might look at the docs for Haskell's Criterion benchmarking library:

  http://www.serpentine.com/criterion/tutorial.html

You presumably won't be able to use it directly (unless you're coding in
Haskell), and lots of the doc is very Haskell-specific, but it can give
you an idea of how a fancy package of this sort goes about things.

[toc] | [prev] | [next] | [standalone]

#24914

Paul Rubin wrote:
> Les Cargill <lcargill99@comcast.com> writes:
>> Obviously a histogram is in order. But wait!.
>> It seems to me that I should be able to estimate a histogram based on:
>> - minimum time
>> - maximum time
>> - average time.
>> What is this sort of ... inquiry called, and how do I stop being quite
>> so dull about it?
>
> You want a probability distribution of the time it takes to do an
> operation?
>
> You might look at the docs for Haskell's Criterion benchmarking library:
>
>    http://www.serpentine.com/criterion/tutorial.html
>
> You presumably won't be able to use it directly (unless you're coding in
> Haskell), and lots of the doc is very Haskell-specific, but it can give
> you an idea of how a fancy package of this sort goes about things.
>

The Haskell reference is quite fruitful. Thanks for that.

The core mechanism of "bench" is a KDE estimator. I'd forgotten
about those. I don't think a min, max and mean are sufficient to
create a KDE estimator off board unless yoo know precisely the number
of modes.

There is a C++ library called libkde. It won't fit, and
I don't have enough bandwidth to use it off-target.

In this case, the better part of valor is to add more and
more fine grained buckets until the buckets become more constant -
mix, max and average are closer together. That
yields a mildly complicated model for each path through the code.

Since more buckets mean mo' problems, I've split the benchmarking into
two types controlled by compiler options. That seems to have done the
trick.

Since the thing is essentially a protocol converter, the four
buckets[1] are enough to make design decisions. I also now
have a plan for estimating dropped "packets".

[1] read from A, read from B, write to A, write to B.

-- 
Les Cargill

[toc] | [prev] | [next] | [standalone]

#24905

On 07/09/16 06:50, Les Cargill wrote:
> I've always suspected there was something clever that I could use for a thing
> like this, but ... I was too afraid to ask about it.
>
> Suppose I am on an extremely deterministic platform.
>
> Suppose I can calculate the number of microsconds it takes to call something[1].
>
> Suppose I can check and calculate things about this count every <n> seconds.
>
> Suppose also that there are two cases for "something" :
>
> - There is nothing to do.
> - There is something to do. "Something" is constrained to "one thing" by design.
>
> Let's also assume I can (cheaply) calculate the minimum and maximum number of
> microseconds to do either of "nothing" or "something".
>
> Obviously a histogram is in order. But wait!.
>
> It seems to me that I should be able to estimate a histogram based on:
> - minimum time
> - maximum time
> - average time.
>
> What is this sort of ... inquiry called, and how do I stop being quite
> so dull about it?
>
> [1] I suspect that one answer is to extend the number of buckets
> into more finely grained measurements of sub-operations of "something".
> Example: a check for "serial port has data" vs. "relieve one unit from serial
> port" vs the overhead from all the dance surrounding those two.

Latencies are usually displayed in the form of a
cumulative distribution function, where the x-axis
is time and the y-axis range is 0 to 1 indicating
the probability that the latency is less than the
time on the x-axis. Effectively the CDF is the
integral of the PDF.

Either the PDF or CDF will show multimodal
distributions and extreme values.

For hard real-time systems either the absolute min
or max is necessary. For soft real-time systems the
5th or 95th percentile is more useful. (or 1/99th
percentile etc).

With the CDF it is trivial to read off the various
percentiles.

[toc] | [prev] | [next] | [standalone]

#24915

Tom Gardner wrote:
> On 07/09/16 06:50, Les Cargill wrote:
>> I've always suspected there was something clever that I could use for
>> a thing
>> like this, but ... I was too afraid to ask about it.
>>
>> Suppose I am on an extremely deterministic platform.
>>
>> Suppose I can calculate the number of microsconds it takes to call
>> something[1].
>>
>> Suppose I can check and calculate things about this count every <n>
>> seconds.
>>
>> Suppose also that there are two cases for "something" :
>>
>> - There is nothing to do.
>> - There is something to do. "Something" is constrained to "one thing"
>> by design.
>>
>> Let's also assume I can (cheaply) calculate the minimum and maximum
>> number of
>> microseconds to do either of "nothing" or "something".
>>
>> Obviously a histogram is in order. But wait!.
>>
>> It seems to me that I should be able to estimate a histogram based on:
>> - minimum time
>> - maximum time
>> - average time.
>>
>> What is this sort of ... inquiry called, and how do I stop being quite
>> so dull about it?
>>
>> [1] I suspect that one answer is to extend the number of buckets
>> into more finely grained measurements of sub-operations of "something".
>> Example: a check for "serial port has data" vs. "relieve one unit from
>> serial
>> port" vs the overhead from all the dance surrounding those two.
>
> Latencies are usually displayed in the form of a
> cumulative distribution function, where the x-axis
> is time and the y-axis range is 0 to 1 indicating
> the probability that the latency is less than the
> time on the x-axis. Effectively the CDF is the
> integral of the PDF.
>
> Either the PDF or CDF will show multimodal
> distributions and extreme values.
>
> For hard real-time systems either the absolute min
> or max is necessary. For soft real-time systems the
> 5th or 95th percentile is more useful. (or 1/99th
> percentile etc).
>
> With the CDF it is trivial to read off the various
> percentiles.


That may be something I try, but for now, even a simple CDF
requires a lot of RAM for a target with 8K main memory:)

But I'll continue playing with it.

-- 
Les Cargill

[toc] | [prev] | [next] | [standalone]

#24919

On 07/09/16 19:10, Les Cargill wrote:
> Tom Gardner wrote:
>> On 07/09/16 06:50, Les Cargill wrote:
>>> I've always suspected there was something clever that I could use for
>>> a thing
>>> like this, but ... I was too afraid to ask about it.
>>>
>>> Suppose I am on an extremely deterministic platform.
>>>
>>> Suppose I can calculate the number of microsconds it takes to call
>>> something[1].
>>>
>>> Suppose I can check and calculate things about this count every <n>
>>> seconds.
>>>
>>> Suppose also that there are two cases for "something" :
>>>
>>> - There is nothing to do.
>>> - There is something to do. "Something" is constrained to "one thing"
>>> by design.
>>>
>>> Let's also assume I can (cheaply) calculate the minimum and maximum
>>> number of
>>> microseconds to do either of "nothing" or "something".
>>>
>>> Obviously a histogram is in order. But wait!.
>>>
>>> It seems to me that I should be able to estimate a histogram based on:
>>> - minimum time
>>> - maximum time
>>> - average time.
>>>
>>> What is this sort of ... inquiry called, and how do I stop being quite
>>> so dull about it?
>>>
>>> [1] I suspect that one answer is to extend the number of buckets
>>> into more finely grained measurements of sub-operations of "something".
>>> Example: a check for "serial port has data" vs. "relieve one unit from
>>> serial
>>> port" vs the overhead from all the dance surrounding those two.
>>
>> Latencies are usually displayed in the form of a
>> cumulative distribution function, where the x-axis
>> is time and the y-axis range is 0 to 1 indicating
>> the probability that the latency is less than the
>> time on the x-axis. Effectively the CDF is the
>> integral of the PDF.
>>
>> Either the PDF or CDF will show multimodal
>> distributions and extreme values.
>>
>> For hard real-time systems either the absolute min
>> or max is necessary. For soft real-time systems the
>> 5th or 95th percentile is more useful. (or 1/99th
>> percentile etc).
>>
>> With the CDF it is trivial to read off the various
>> percentiles.
>
>
> That may be something I try, but for now, even a simple CDF
> requires a lot of RAM for a target with 8K main memory:)
>
> But I'll continue playing with it.

You store the PDF, and compute the CDF when needed.
The range and resolution translates directly into
the number of "buckets", and the number of bits in
each bucket is determined by the number of samples.
No surprises there.

Sometimes pulsing an output line high during the
interval, and averaging with an RC filter is
sufficient.

[toc] | [prev] | [standalone]

Back to top | Article view | comp.arch.embedded

csiph-web