veer66

(Posted on 2022-04-20)

According to this thread, I compared using only lazy sequences with transducers.

To add the I/O factor, I prepared a data file called “fake.txt” using the program below:

(with-open [w (io/writer "fake.txt")]
  (doseq [n (range 10000000)]
    (.write w (str n "\n"))))

F1 is the lazy-sequence-based version. It reads data from “fake.txt” and does a few steps of computations.

(defn f1
  []
  (with-open [r (io/reader "fake.txt")]
    (->> (line-seq r)
         (map parse-long)
         (map inc)
         (filter even?)
         (map inc)
         (reduce + 0))))

F2 is the transducer-based version of F1.

(defn f2
  []
  (with-open [r (io/reader "fake.txt")]
    (transduce (comp (map parse-long)
                     (map inc)
                     (filter even?)
                     (map inc))
               +
               (line-seq r))))

I evaluated them using Criterium.

(with-progress-reporting (quick-bench (f1) :verbose))
(with-progress-reporting (quick-bench (f2) :verbose))

Here is the result.

#################### F1 ###################
Evaluation count : 6 in 6 samples of 1 calls.
Execution time sample mean : 3.811858 sec
Execution time mean : 3.812064 sec

#################### F2 ###################
Evaluation count : 6 in 6 samples of 1 calls.
Execution time sample mean : 1.490624 sec
Execution time mean : 1.490777 sec

F1, which is the lazy sequence version, took 3.812064 seconds. F2, which is the transducer version, took 1.490777. So the transducer version is 155.71% faster than the lazy sequence version.

In brief, this biased experiment shows the transducer version is much faster than the pure lazy sequence version.