Quick Hit: Comparison of “Whole File Reading” Methods

by hrbrmstr on August 7, 2020

(This is part 1 of n posts using this same data; n will likely be 2-3, and the posts are more around optimization than anything else.)

I recently had to analyze HTTP response headers (generated by a HEAD request) from around 74,000 sites (each response stored in a text file). They look like this:

HTTP/1.1 200 OKDate: Mon, 08 Jun 2020 14:40:45 GMTServer: ApacheLast-Modified: Sun, 26 Apr 2020 00:06:47 GMTETag: "ace-ec1a0-5a4265fd413c0"Accept-Ranges: bytesContent-Length: 967072X-Frame-Options: SAMEORIGINContent-Type: application/x-msdownload

I do this quite a bit in R when we create new studies at work, but I’m usually only working with a few files. In this case I had to go through all these files to determine if a condition hypothesis (more on that in one of the future posts) was accurate.

Reading in a bunch of files (each one into a string) is fairly straightforward in R since readChar() will do the work of reading and we just wrap that in an iterator:

length(fils)## [1] 73514 # check file size distributionsummary(  vapply(    X = fils,    FUN = file.size,    FUN.VALUE = numeric(1),    USE.NAMES = FALSE  ))## Min. 1st Qu.  Median    Mean 3rd Qu.    Max. ## 19.0   266.0   297.0   294.8   330.0  1330.0 # they're all super smallsystem.time(  vapply(    X = fils,     FUN = function(.f) readChar(.f, file.size(.f)),     FUN.VALUE = character(1),     USE.NAMES = FALSE  ) -> tmp )##  user  system elapsed ## 2.754   1.716   4.475

NOTE: You can use lapply() or sapply() to equal effect as they all come in around 5 seconds on a modern SSD-backed system.

Sponsorships Available

Now, five seconds is completely acceptable (though that brief pause does feel awfully slow for some reason), but can we do better? I mean we do have some choices when it comes to slurping up the contents of a file into a length 1 character vector:

base::readChar()
readr::read_file()
stringi::stri_read_raw() (+ rawToChar())

Do any of them beat {base}? Let’s see (using the largest of the files):

library(stringi)library(readr)library(microbenchmark)largest <- fils[which.max(sapply(fils, file.size))]file.size(largest)## [1] 1330microbenchmark(  base = readChar(largest, file.size(largest)),  readr = read_file(largest),  stringi = rawToChar(stri_read_raw(largest)),  times = 1000,  control = list(warmup = 100))## Unit: microseconds##     expr     min       lq      mean   median       uq     max neval##     base  79.862  93.5040  98.02751  95.3840 105.0125 161.566  1000##    readr 163.874 186.3145 190.49073 189.1825 192.1675 421.256  1000##  stringi  52.113  60.9690  67.17392  64.4185  74.9895 249.427  1000

I had predicted that the {stringi} approach would be slower given that we have to explicitly turn the raw vector into a character vector, but it is modestly faster. ({readr} has quite a bit of functionality baked into it — for good reasons — which doesn’t help it win any performance contests).

I still felt there had to be an even faster method, especially since I knew that the files all had HTTP response headers and that they every one of them could each be easily read into a string in (pretty much) one file read operation. That knowledge will let us make a C++ function that cuts some corners (more like “sands” some corners, really). We’ll do that right in R via {Rcpp} in this function (annotated in C++ code comments):

library(Rcpp)cppFunction(code = 'String cpp_read_file(std::string fil) {  // our input stream  std::ifstream in(fil, std::ios::in | std::ios::binary);  if (in) { // we can work with the file  #ifdef Win32    struct _stati64 st; // gosh i hate windows    _wstati64(wfn, &st) // this shld work but I did not test it  #else    struct stat st;    stat(fil.c_str(), &st);  #endif    std::string out; // where we will store the contents of the file    out.reserve(st.st_size); // make string size == file size    in.seekg(0, std::ios::beg); // ensure we are at the beginning    in.read(&out[0], out.size()); // read in the file    in.close();    return(out);  } else {    return(NA_STRING); // file missing or other errors returns NA  }}', includes = c(  "#include <fstream>",  "#include <string>",  "#include <sys/stat.h>"))

Is that going to be faster?

microbenchmark(  base = readChar(largest, file.size(largest)),  readr = read_file(largest),  stringi = rawToChar(stri_read_raw(largest)),  rcpp = cpp_read_file(largest),  times = 1000,  control = list(warmup = 100))## Unit: microseconds##     expr     min       lq      mean   median       uq     max neval##     base  80.500  91.6910  96.82752  94.3475 100.6945 295.025  1000##    readr 161.679 175.6110 185.65644 186.7620 189.7930 399.850  1000##  stringi  51.959  60.8115  66.24508  63.9250  71.0765 171.644  1000##     rcpp  15.072  18.3485  21.20275  21.0930  22.6360  62.988  1000

It sure looks like it, but let’s put it to the test:

system.time(  vapply(    X = fils,     FUN = cpp_read_file,     FUN.VALUE = character(1),     USE.NAMES = FALSE  ) -> tmp )##  user  system elapsed ## 0.446   1.244   1.693

I’ll take a two-second wait over a five-second wait any day!

FIN

I have a few more cases coming up where there will be 3-5x the number of (similar) files that I’ll need to process, and this optimization will shave time off as I iterate through each analysis, so the trivial benefits here will pay off more down the road.

The next post in this particular series will show how to use the {future} family to reduce the time it takes to turn those HTTP headers into data we can use.

If I missed your favorite file slurping function, drop a note in the comments and I’ll update the post with new benchmarks.