Optimize the runtime: change the weight of edges in an igraph takes long time. Is there a way to optimize it?

Question

I am searching for a set of edges in an igraph built from an osmar object and would like to change the weight of these. Since my graph is quite big, this task takes quite a long time. Since I run this function in a loop the runtime grows even bigger.

Is there a way I could optimize this?

Here is the code:

library(osmar)
library(igraph)
library(tidyr)
library(dplyr)

### Get data ----
src <- osmsource_api(url = "https://api.openstreetmap.org/api/0.6/")
muc_bbox <- center_bbox(11.575278, 48.137222, 1000, 1000)
muc <- get_osm(muc_bbox, src)

### Reduce to highways: ----
hways <- subset(muc, way_ids = find(muc, way(tags(k == "highway"))))
hways <- find(hways, way(tags(k == "name")))
hways <- find_down(muc, way(hways))
hways <- subset(muc, ids = hways)

#### Plot data ----
## Plot complete data and highways on top:
plot(muc)
plot_ways(muc, col = "lightgrey")
plot_ways(hways, col = "coral", add = TRUE)

### Define route start and end nodes: ----
id<-find(muc, node(tags(v %agrep% "Sendlinger Tor")))[1]
hway_start_node <-find_nearest_node(muc, id, way(tags(k == "highway"))) 
hway_start <- subset(muc, node(hway_start_node))

id <- find(muc, node(attrs(lon > 11.58 & lat > 48.15)))[1]
hway_end_node <- find_nearest_node(muc, id, way(tags(k == "highway")))
hway_end <- subset(muc, node(hway_end_node))

## Add the route start and and nodes to the plot:
plot_nodes(hway_start, add = TRUE, col = "red", pch = 19, cex = 2)
plot_nodes(hway_end, add = TRUE, col = "red", pch = 19, cex = 2)

### Create street graph ----
gr <- as.undirected(as_igraph(hways))

### Compute shortest route: ----
# Calculate route
route <- function(start_node,end_node) {
  get.shortest.paths(gr,
                     from = as.character(start_node),
                     to = as.character(end_node), 
                     mode = "all")[[1]][[1]]}
# Plot route
plot.route <- function(r,color) {
  r.nodes.names <- as.numeric(V(gr)[r]$name)
  r.ways <- subset(hways, ids = osmar::find_up(hways, node(r.nodes.names)))
  plot_ways(r.ways, add = TRUE, col = color, lwd = 2)
}
nways <-  1
numway <- 1
r <- route(hway_start_node,hway_end_node)

# Plot route

color <- colorRampPalette(c("springgreen","royalblue"))(nways)[numway]
plot.route(r,color)


## Route details ----
# Construct a new osmar object containing only elements 
# related to the nodes defining the route:
route_nodes <- as.numeric(V(gr)[r]$name)
route_ids <- find_up(hways, node(route_nodes))

osmar.route <- subset(hways, ids = route_ids)
osmar.nodes.ids <- osmar.route$nodes$attrs$id

# Extract the nodes’ coordinates,
osmar.nodes.coords <- osmar.route$nodes$attrs[, c("lon", "lat")]
osmar.nodes <- cbind(data.frame(ids = osmar.nodes.ids),
                     data.frame(ids_igraph = as.numeric(V(gr)[r]) ),
                     osmar.nodes.coords) 


## Find edges ids containing points of interest ----
wished.coords <- data.frame(wlon = c(11.57631),
                            wlat = c(48.14016)) 


# Calculate all distances
distances <- crossing(osmar.nodes,wished.coords) %>%
             mutate(dist = geosphere::distHaversine(cbind(lon,lat),cbind(wlon,wlat)))


# Select nodes below maximum distance :
mindist <- 50 #m

wished.nodes <- distances %>% filter(dist < mindist)

# Select edges incident to these nodes :
selected.edges <- unlist(incident_edges(gr,V(gr)[wished.nodes$ids_igraph]))

This is where the slowdown occurs: Weight of selected edges, change it by multiplying it with 10
E(gr)[selected.edges]$weight<-E(gr)[selected.edges]$weight*10

This is where the slowdown occurs: Weight of selected edges, change it by multiplying it with 10

E(gr)[selected.edges]$weight<-E(gr)[selected.edges]$weight*10

Maybe I could use a hashmap?

UPDATE

hashmap

Unit: seconds

Hashmap:

expr           min       lq     mean   median      uq      max     neval 

Hashmap      3.248543 3.289474 3.472038 3.324417 3.734050 4.188924   100 

Without      3.267549 3.333012 3.557179 3.367015 3.776429 5.643784   100

Sadly it does not seemt to bring a lot of improvement.


library(hashmap) 
#https://github.com/nathan-russell/hashmap
         H <- hashmap(E(gr)[selected.edges],E(gr)[selected.edges]$weight)
         sapply(H$find(E(grr)[selected.edges]), function(x) x * 10)

UPDATE: According to igraph doc, igraph is thread safe so I could use parallel.

Im currently trying this:

no_cores <- detectCores(logical = FALSE) 
 data <- split(selected.edges,factor(sort(rank(selected.edges)%%no_cores)))
 c_result <- mclapply(1:no_cores, function(x) {
 E(gr)[unlist(data[[x]])]$weight * 1000 / mean_value }, mc.cores = no_cores) 
   
     E(gr)[unlist(data)]$weight<-unlist(c_result)

I wonder why do I have to do the "writing step" outside of parallel loop. As I was trying to write the weight back to igraph within the loop it did not work ie weight did not get updated.

Thank you in advance! BR

Answer 1

As demonstrated in Advanced R, implementation performance in R can greatly vary depending on the syntax.

E(gr)[selected.edges]$weight<-E(gr)[selected.edges]$weight*10

is a valid syntax, but it can also be formulated otherwise:

set.edge.attribute(graph=gr,name="weight",index=selected.edges,value=10*get.edge.attribute(graph=gr,name="weight",index=selected.edges))

So let's compare both solutions :

microbenchmark::microbenchmark(
  ref={E(gr)[selected.edges]$weight<-E(gr)[selected.edges]$weight*10},
  new={set.edge.attribute(graph=gr,name="weight",index=selected.edges,value=10*get.edge.attribute(graph=gr,name="weight",index=selected.edges))})

Unit: microseconds
 expr       min        lq       mean    median        uq       max neval cld
  ref 15920.404 16567.788 17793.4412 17111.583 18491.685 25867.477   100   b
  new   246.974   266.462   296.5088   278.769   292.718   662.974   100  a 

@Andreas, can you please check on a bigger dataset if this could be a solution to your problem?