IntenseFinance/nosqlbench
3
0
Fork 0
mirror of https://github.com/nosqlbench/nosqlbench.git synced 2025-02-25 18:55:28 -06:00
Code Issues Packages Projects Releases Wiki Activity

improve distributions docs

Browse Source
This commit is contained in:
Jonathan Shook 2020-05-01 13:36:22 -05:00
parent 088e9314fa
commit 77ca72a7fe
1 changed files with 44 additions and 53 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

97
virtdata-userlibs/src/main/resources/docs/category_blurbs/funcref_distributions.md
View File

@ -12,81 +12,72 @@ These distributions break down into two main categories:
### Continuous Distributions
These are distributions over real numbers like 23.4323, with
continuity across the values. Each of the continuous distributions can
provide samples that fall on an interval of the real number line.
Continuous probability distributions include the *Normal* distribution,
and the *Exponential* distribution, among many others.
These are distributions over real numbers like 23.4323, with continuity across the values. Each of the continuous
distributions can provide samples that fall on an interval of the real number line. Continuous probability distributions
include the *Normal* distribution, and the *Exponential* distribution, among many others.
### Discrete Distributions
Discrete distributions, also known as *integer distributions* have only
whole-number valued samples. These distributions include the *Binomial*
distribution, the *Zipf* distribution, and the *Poisson* distribution,
among others.
Discrete distributions, also known as *integer distributions* have only whole-number valued samples. These distributions
include the *Binomial* distribution, the *Zipf* distribution, and the *Poisson* distribution, among others.
## Hashed or Mapped
### hashed samples
Generally, you will want to "randomly sample" from a probability distribution.
This is handled automatically by the functions below if you do not override the
defaults. **The `hash` mode is the default sampling mode for probability
distributions.** This is accomplished by computing an internal on the unit
interval variate input before using the resulting value to map into the sampling
curve. This is called the `hash` sampling mode by VirtData. You can put `hash`
into the modifiers as explained below if you want to document it explicitly.
Generally, you will want to "randomly sample" from a probability distribution. This is handled automatically by the
functions below if you do not override the defaults. **The `hash` mode is the default sampling mode for probability
distributions.** This is accomplished by hashing the input before using the resulting value with the sampling curve.
This is called the `hash` sampling mode by VirtData. You can put `hash` into the modifiers as explained below if you
want to document it explicitly.
### mapped samples
The method used to sample from these distributions depends on a mathematical
function called the cumulative probability function, or more specifically
the inverse of it. Having this function computed over some interval allows
one to sample the shape of a distribution progressively if desired. In
other words, it allows for some *percentile-like* view of values within
a given probability distribution. This mode of using the inverse cumulative
density function is known as the `map` mode in VirtData, as it allows one
to map a unit interval variate in a deterministic way to a density
sampling curve. To enable this mode, simply pass `map` as one of the
function modifiers for any function in this category.
The method used to sample from these distributions depends on a mathematical function called the cumulative probability
density function, or more specifically the inverse of it. Having this function computed over some interval allows one to
sample the shape of a distribution progressively if desired. In other words, it allows for some *percentile-like* view
of values within a given probability distribution. This mode of using the inverse cumulative density function is known
as the `map` mode in VirtData, as it allows one to map a unit interval variate in a deterministic way to a density
sampling curve. To enable this mode, simply pass `map` as one of the function modifiers for any function in this
category.
## Interpolated or Computed Samples
When sampling from mathematical models of probability densities, performance
between different densities can vary drastically. This means that you may
end up perturbing the results of your test in an unexpected way simply
by changing parameters of your testing distributions. Even worse, some
densities have painful corner cases in performance, like 'Zipf', which
can make tests unbearably slow and flawed as they chew up CPU resources.
When sampling from mathematical models of probability densities, performance between different densities can vary
drastically. This means that you may end up perturbing the results of your test in an unexpected way simply by changing
parameters of your testing distributions. Even worse, some densities have painful corner cases in performance, like
'Zipf', which can make tests unbearably slow and flawed as they chew up CPU resources.
::: info
Functions like 'Zipf' can still take a long time to initialize for certain parameters. If you are seeing a workload that
seems to hang while initializing, it might be computing complex integrals for large parameters of Zipf. We hope to
pre-compute and cache these at a future time to avoid this type of impact. For now, just be aware that some parameters
on some density curves can be expensive to compute _even during initialization_.
:::
### Interpolated Samples
For this reason, interpolation is built-in to these sampling functions.
**The default mode is `interpolate`.** This means that the sampling
function is pre-computed over 1000 equidistant points in the unit interval,
and the result is shared among all threads as a look-up-table for
interpolation. This makes all statistical sampling functions perform nearly
identically at runtime (after initialization, a one time cost).
This does have the minor side effect of a little loss in accuracy, but
the difference is generally negligible for nearly all performance testing
cases.
For this reason, interpolation is built-in to these sampling functions. **The default mode is `interpolate`.** This
means that the sampling function is pre-computed over 1000 equidistant points in the unit interval (0.0,1.0), and the
result is shared among all threads as a look-up-table for interpolation. This makes all statistical sampling functions
perform nearly identically at runtime (after initialization, a one time cost). This does have the minor side effect of a
little loss in accuracy, but the difference is generally negligible for nearly all performance testing cases.
### Computed Samples
Conversely, `compute` mode sampling calls the sampling function every
time a sample is needed. This affords a little more accuracy, but is generally
not preferable to the default interpolated mode. You'll know if you need
computed samples. Otherwise, it's best to stick with interpolation so that
you spend more time testing your target system and less time testing
your data generation functions.
Conversely, `compute` mode sampling calls the sampling function every time a sample is needed. This affords a little
more accuracy, but is generally *not* preferable to the default interpolated mode. You'll know if you need computed
samples. Otherwise, it's best to stick with interpolation so that you spend more time testing your target system and
less time testing your data generation functions.
## Input Range
All of these functions take a long as the input value for sampling. This
is similar to how the unit interval (0.0,1.0) is used in mathematics
and statistics, but more tailored to modern system capabilities. Instead
of using the unit interval, we simply use the interval of all positive
longs. This provides more compatibility with other functions in VirtData,
including hashing functions.
All of these functions take a long as the input value for sampling. This is similar to how the unit interval (0.0,1.0)
is used in mathematics and statistics, but more tailored to modern system capabilities. Instead of using the unit
interval, we simply use the interval of all positive longs. This provides more compatibility with other functions in
VirtData, including hashing functions. Internally, this value is automatically converted to a unit interval variate as
needed to work well with the distributions from Apache Math.