---
title: 'Speed improvement'
author: "Thijs Janzen"
date: "`r Sys.Date()`"
output: rmarkdown::html_vignette
vignette: >
%\VignetteEncoding{UTF-8}
%\VignetteIndexEntry{Speed improvement}
%\VignetteEngine{knitr::rmarkdown}
---
```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = FALSE)
knitr::opts_chunk$set(fig.width = 6)
knitr::opts_chunk$set(fig.height = 6)
speed_data <- read.table("https://raw.githubusercontent.com/thijsjanzen/treestats-scripts/main/Figure_S3_S4/timings.txt", header = TRUE) # nolint
```
## Speed
An important goal during the development of treestats, has been to
provide not only a vast collection of phylogenetic tree statistics, but
also ensure that calculation of these statistics is fast and reliable.
During development, results of the developed code have continuously been
tested against reference code from other packages. After ensuring
correctness, profiling methods have been used to improve calculation
speed.
### Timing the relationship with tree size
Tree size is an important factor in determining calculation performance.
We have varied tree size from 10 to 1000 in logarithmicly appropriate
steps, calculating statistics for 10 randomly generated Yule trees per
tree size.
```{r out.width="100%", echo=FALSE}
knitr::include_graphics("https://github.com/thijsjanzen/treestats/blob/main/layout/Figure_S3.png?raw=true") # nolint
```
Results show that for many statistics, treestats code provides a serious speed
improvement over the old R code. Furthermore, scaling with tree size is much
better (consider for instance mpd and mntd, which scale exponentially with tree
size using R, but linear with using treestats). Two functions are faster in
other packages: colless and gamma. These two functions are slower in treestats
due to additional checks before performing the calculations, these checks verify
whether the tree is binary (for colless) or ultrametric (for the gamma statistic).
### Worst culprits
There are three functions that require a considerably larger amount of time to
be calculated on a tree of 1000 tips, compared to all other functions (see
Figure below). These three functions rely on calculation of the Eigen values of
a (Laplacian) matrix. Unfortunately, calculating Eigen values is a costly
endeavour, and can not be further optimized in-house - we use the base Eigen
functions for that already.
```{r summarise_data, echo=FALSE}
res <- c()
for (x in unique(speed_data$treestatsfunction)) {
a <- subset(speed_data, speed_data$treestatsfunction == x &
speed_data$ntips == 1000 &
speed_data$method == "treestats")
res <- rbind(res, c(x, mean(a$time)))
}
res <- data.frame("treestats_function" = res[, 1],
"time" = res[, 2])
res$time <- as.numeric(res$time)
res2 <- res[order(res$time), ]
```
```{r out.width="100%", echo = FALSE}
opar <- par(no.readonly = TRUE)
par(mar = c(8, 4, 4, 4))
barplot(res2$time, names.arg = res2$treestats_function, las = 2,
cex.names = 0.4, ylab = "Time", log = "y")
par(opar)
```
### Speed improvement
So how much does the speed improve for large trees? We see that for many statistics,
the speed improvement is at least a factor 10, with for some statistics (mntd, mpd,
psv) a speed improvement around four orders of magnitude. Currently, for many
functions speed is limited by preliminary checks if the used tree is ultrametric
and/or binary (if required), and code converting a phy object to C++ objects.
Nevertheless, improvements are considerable and we hope that these improvements
in calculation speed allow for a broader application of many summary statistics.
```{r out.width="100%", echo=FALSE}
knitr::include_graphics("https://github.com/thijsjanzen/treestats/blob/main/layout/Figure_S4.png?raw=true") # nolint
```