Compiling JMeter runs

JMeter CSV Log Format

The JMeter format was described in the previous JMeter article. In the readLog function an additional column called version ws added.

Aggregate

tr$summary <-
  tr$log %>%
  select(label, version, elapsed) %>%
  mutate(version = as.factor(version)) %>%
  group_by(version,label) %>%
  summarise(average=mean(elapsed)) %>%
  arrange(desc(average))

Identify slowest pages

For the final plot, ggplot will require the data in long format. Also I also want to order the samples by page so i can quickly see the improvement or detriment across for a page in the different versions.

I want to see the slowest pages regardless of version.

To identify the slow pages that will be shown in the graph, I will collect the data into wide form giving me one row for each page, add a column “slowest” with the slowest time for the page across versions, then sort and filter on this new value. To convert long to wide the function spread() is supplied by dplyr.

Limit to the 6 slowest pages

# For each page identify the slowest result across all versions.
tr$wide <- tr$summary %>% 
    pivot_wider(names_from = version, values_from = average) %>% 
    mutate(label, slowest = select(., levels(tr$summary$version)) %>% reduce(pmax)) 
    
# Add the slowest metric to the long form data
tr$osummary <- inner_join(tr$summary, select(tr$wide, label, slowest) ) %>%
  arrange(desc(slowest), label, version) %>%
  head(6 * length(unique(tr$summary$version))) %>%
  mutate(label = as.factor(label))

# Reverse the versions factor for the desired result in the plot.
tr$osummary$version <- factor(tr$osummary$version, levels = rev(levels(tr$osummary$version)))  

Plot

positions <- rev(tr$osummary$label)
ggplot(data = tr$osummary,  aes(x=label, y=average, fill=version)) +
  geom_bar(stat="identity", position = "dodge", 
           width= length(levels(tr$summary$version))*0.75 ) + 
  scale_x_discrete(limits = positions) +
  coord_flip()  +
  scale_fill_discrete(guide = guide_legend(reverse=TRUE)) +
  theme_bw()

Table Breakdown

tr$v$oldest <- first(rev(levels(tr$summary$version)))
tr$v$newest <- last(rev(levels(tr$summary$version)))

tr$wide %>%
  mutate(improvement =select(., c(tr$v$oldest, tr$v$newest)) %>% reduce(function(old,new) scales::percent((new-old)/new))) %>%
  select(-slowest) %>%
 knitr::kable()

Observations

The plot and table indicate that slow pages, anything over 1/4 of a second has been treated with significant succes. The search page is predictably the slowest but is working with respectable performance. Any change to the fastest pages is smaller than the number of significant digits used in the table formatting.