如何提高Clojure中的文本处理性能？

Question

我正在Clojure中编写一个简单的桌面搜索引擎,以此来了解有关该语言的更多信息.到目前为止,我的程序文本处理阶段的表现非常糟糕.

在文本处理期间,我要:

• 清理不需要的字符;
• 将字符串转换为小写;
• 拆分文档以获取单词列表;
• 构建一个地图,将每个单词与其在文档中的出现相关联.

这是代码:

(ns txt-processing.core
  (:require [clojure.java.io :as cjio])
  (:require [clojure.string :as cjstr])
  (:gen-class))

(defn all-files [path]
  (let [entries (file-seq (cjio/file path))]
    (filter (memfn isFile) entries)))

(def char-val
  (let [value #(Character/getNumericValue %)]
    {:a (value \a) :z (value \z)
     :A (value \A) :Z (value \Z)
     :0 (value \0) :9 (value \9)}))

(defn is-ascii-alpha-num [c]
  (let [n (Character/getNumericValue c)]
    (or (and (>= n (char-val :a)) (<= n (char-val :z)))
        (and (>= n (char-val :A)) (<= n (char-val :Z)))
        (and (>= n (char-val :0)) (<= n (char-val :9))))))

(defn is-valid [c]
    (or (is-ascii-alpha-num c)
        (Character/isSpaceChar c)
        (.equals (str \newline) (str c))))

(defn lower-and-replace [c]
  (if (.equals (str \newline) (str c)) \space (Character/toLowerCase c)))

(defn tokenize [content]
  (let [filtered (filter is-valid content)
        lowered (map lower-and-replace filtered)]
    (cjstr/split (apply str lowered) #"\s+")))

(defn process-content [content]
  (let [words (tokenize content)]
    (loop [ws words i 0 hmap (hash-map)]
      (if (empty? ws)
        hmap
        (recur (rest ws) (+ i 1) (update-in hmap [(first ws)] #(conj % i)))))))

(defn -main [& args]
  (doseq [file (all-files (first args))]
    (let [content (slurp file)
          oc-list (process-content content)]
      (println "File:" (.getPath file)
               "| Words to be indexed:" (count oc-list )))))

由于我在Haskell中有另一个实现此问题的方法,因此我在以下输出中进行了比较.

Clojure版本:

$ lein uberjar
Compiling txt-processing.core
Created /home/luisgabriel/projects/txt-processing/clojure/target/txt-processing-0.1.0-SNAPSHOT.jar
Including txt-processing-0.1.0-SNAPSHOT.jar
Including clojure-1.5.1.jar
Created /home/luisgabriel/projects/txt-processing/clojure/target/txt-processing-0.1.0-SNAPSHOT-standalone.jar
$ time java -jar target/txt-processing-0.1.0-SNAPSHOT-standalone.jar ../data
File: ../data/The.Rat.Racket.by.David.Henry.Keller.txt | Words to be indexed: 2033
File: ../data/Beyond.Pandora.by.Robert.J.Martin.txt | Words to be indexed: 1028
File: ../data/Bat.Wing.by.Sax.Rohmer.txt | Words to be indexed: 7562
File: ../data/Operation.Outer.Space.by.Murray.Leinster.txt | Words to be indexed: 7754
File: ../data/The.Reign.of.Mary.Tudor.by.James.Anthony.Froude.txt | Words to be indexed: 15418
File: ../data/.directory | Words to be indexed: 3
File: ../data/Home.Life.in.Colonial.Days.by.Alice.Morse.Earle.txt | Words to be indexed: 12191
File: ../data/The.Dark.Door.by.Alan.Edward.Nourse.txt | Words to be indexed: 2378
File: ../data/Storm.Over.Warlock.by.Andre.Norton.txt | Words to be indexed: 7451
File: ../data/A.Brief.History.of.the.United.States.by.John.Bach.McMaster.txt | Words to be indexed: 11049
File: ../data/The.Jesuits.in.North.America.in.the.Seventeenth.Century.by.Francis.Parkman.txt | Words to be indexed: 14721
File: ../data/Queen.Victoria.by.Lytton.Strachey.txt | Words to be indexed: 10494
File: ../data/Crime.and.Punishment.by.Fyodor.Dostoyevsky.txt | Words to be indexed: 10642

real    2m2.164s
user    2m3.868s
sys     0m0.978s

Haskell版本:

$ ghc -rtsopts --make txt-processing.hs 
[1 of 1] Compiling Main             ( txt-processing.hs, txt-processing.o )
Linking txt-processing ...
$ time ./txt-processing ../data/ +RTS -K12m
File: ../data/The.Rat.Racket.by.David.Henry.Keller.txt | Words to be indexed: 2033
File: ../data/Beyond.Pandora.by.Robert.J.Martin.txt | Words to be indexed: 1028
File: ../data/Bat.Wing.by.Sax.Rohmer.txt | Words to be indexed: 7562
File: ../data/Operation.Outer.Space.by.Murray.Leinster.txt | Words to be indexed: 7754
File: ../data/The.Reign.of.Mary.Tudor.by.James.Anthony.Froude.txt | Words to be indexed: 15418
File: ../data/.directory | Words to be indexed: 3
File: ../data/Home.Life.in.Colonial.Days.by.Alice.Morse.Earle.txt | Words to be indexed: 12191
File: ../data/The.Dark.Door.by.Alan.Edward.Nourse.txt | Words to be indexed: 2378
File: ../data/Storm.Over.Warlock.by.Andre.Norton.txt | Words to be indexed: 7451
File: ../data/A.Brief.History.of.the.United.States.by.John.Bach.McMaster.txt | Words to be indexed: 11049
File: ../data/The.Jesuits.in.North.America.in.the.Seventeenth.Century.by.Francis.Parkman.txt | Words to be indexed: 14721
File: ../data/Queen.Victoria.by.Lytton.Strachey.txt | Words to be indexed: 10494
File: ../data/Crime.and.Punishment.by.Fyodor.Dostoyevsky.txt | Words to be indexed: 10642

real    0m9.086s
user    0m8.591s
sys     0m0.463s

我认为Clojure实现中的(字符串 - > 延迟序列)转换会破坏性能.我怎样才能改进它？

PS:这些测试中使用的所有代码和数据都可以在这里下载.

Answer 1

您可以做的一些事情可能会加快此代码的速度：

1）无需将您的映射chars到char-val，只需在字符之间进行直接值比较即可。其速度更快的原因与 Java 中速度更快的原因相同。

2) 您重复使用str将单字符值转换为完整的字符串。再次考虑直接使用字符值。同样，对象创建速度很慢，与 Java 中一样。

3）您应该替换process-content为clojure.core/frequencies. 也许检查frequencies源代码看看它是如何更快的。

4) 如果必须(hash-map)在循环中更新 a，请使用transient. 请参阅： http: //clojuredocs.org/clojure_core/clojure.core/transient

另请注意，(hash-map)返回 a PersistentArrayMap，因此您每次调用都会创建新实例update-in- 因此速度很慢，以及为什么您应该使用瞬态。

5) 这是你的朋友：(set! *warn-on-reflection* true)- 你有相当多的反思可以从类型提示中受益

 Reflection warning, scratch.clj:10:13 - call to isFile can't be resolved.
 Reflection warning, scratch.clj:13:16 - call to getNumericValue can't be resolved.
 Reflection warning, scratch.clj:19:11 - call to getNumericValue can't be resolved.
 Reflection warning, scratch.clj:26:9 - call to isSpaceChar can't be resolved.
 Reflection warning, scratch.clj:30:47 - call to toLowerCase can't be resolved.
 Reflection warning, scratch.clj:48:24 - reference to field getPath can't be resolved.
 Reflection warning, scratch.clj:48:24 - reference to field getPath can't be resolved.