Table of Contents

• Benchmarks Results
  • Test Methods
    • Running tests
    • Generating results
    • Benchmarks framework
    • Profiles
  • Results
    • get
      • Details
      • Results
    • assoc
      • Details
      • Results
      • By keys
      • By width
    • merge
      • Details
        • Fast map merge
        • Inline Merge
      • Results
        • Fast Map merge
        • Inline merge & fast merge
      • By width
      • By keys
    • get-in
      • Test details
      • Results
      • By keys
      • By width
    • memoize
      • Test details
      • Results
      • By number of arguments
      • By type of arguments
    • assoc-in
      • Test details
      • Results
      • By keys
      • By width
    • update-in
      • By keys
      • By width
    • select-keys
      • Test details
      • Results
      • By keys
      • By width

Benchmarks Results

Test Methods

Running tests

lein uberjar
java -Xmx8G -Xms8G -XX:+UseG1GC \
  -jar target/uberjar/clj-fast-0.0.2-alpha-standalone.jar \
  --max-width 5 \
  --max-depth 5 \
  --types "keyword?" \
  --name all \
  get get-rec get-in assoc assoc-rec merge select-keys assoc-in update-in

Generating results

Load analysis.clj and:

(ns clj-fast.analysis)
(def raw-data
  (-> "./benchmarks/all-clj-fast-bench.edn"
      load-results
      (update :get-rec #(map (fn [m] (assoc m :width 0)) %))
      (update :merge #(remove (comp #{1} :keys) %))))

(def all-charts
  (conj
   (common-charts raw-data)
   (chart-get :get raw-data)
   (chart-get :get-rec raw-data)
   (chart-assoc-rec raw-data)))

(logify :merge all-charts)

(write-charts all-charts)

Benchmarks framework

Criterium is used to run quick benchmarks.

Profiles

Benchmarks were run with 8GB heap and G1 garbage collection.

Results

get

Details

get was tested on map, record and fast-map, fast-get was tested on fast-map.

Moreover, different get methods were tested:

• map on keyword.
• keyword on map.
• keyword on record.
• .get from record.
• .field from record.

Results

• get from record ~50% slower than from map.
• fast-get from fast-map ~ 2x faster than getting from regular map, with gains increasing as the map becomes larger. (Metosin)
• invoking a map, calling a keyword on a map, accessing valAt all have approximately the same performance, however, there are more levels of indirection when calling a keyword, less than when invoking a map, and zero when calling valAt.
• All other methods besides get are faster
• The fastest way to get a field from a record is a field accessor.

get from map	get from record

assoc

Details

Assoc and fast assoc performance are tested with maps and records.

Results

• assoc to record is as fast as associng to map.
• fast-assoc ~ 5.7% faster than assoc. (Metosin)

By keys

By width

merge

Details

Fast map merge

fast map merge was implemented by Metosin and uses kv-reduce to assoc one map into another. Compare to regular merge for the keys=2 case.

Inline Merge

Several inline implementations were compared:

• inline-merge: uses conj directly on all input maps instead of reducing.
• inline-fast-map-merge: inline merges maps using Metosin's fast-map-merge.
• inline-tmerge: Uses transients to merge the maps. Basic implementation by Joinr.

Results

Fast Map merge

fast-map-merge is faster than regular merge, especially for smaller maps, with diminishing returns as maps get bigger.

Inline merge & fast merge

Sees diminishing returns on the benefit of merging more maps, but some speedup is measurable. tmerge is slower for small maps but for larger maps is faster than regular merge. The speedup by fast-map-merge is about 15-25%, width diminishing returns the bigger maps get.

Execution time is presented in logarithmic scale due to the huge differences for different map sizes.

By width

By keys

get-in

Test details

get-in was tested against an inlined implementation.

Results

Inline implementation faster by a factor of 4-5.

By keys

By width

memoize

Closely related to get-in

Test details

memoize was tested against an inlining implementation using a clojure atom (memoize-n) and one using a concurrent-hash-map (memoize-c)

Results

Different implementation are faster depending on the type of the memoized arguments:

• memoize-n is better when all arguments are keywords or symbols.
• memoize-c is better for any other case.

By number of arguments

By type of arguments

assoc-in

Test details

Assoc-in is tested vs. an inlined implementation with vanilla maps, gets and assoc, all core functions.

Results

• the inlined implementation is always faster and exhibits compounding returns for deeper maps, about ~25ns/key.

By keys

By width

update-in

By keys

By width

select-keys

Test details

select-keys was tested against an inlined implementation.

Results

Inline implementation faster by a factor of 10 or more, depends on the number of selected keys.

By keys

By width