Clojure

clojure.core/*rand* for seedable randomness

Details

  • Type: Enhancement Enhancement
  • Status: Open Open
  • Priority: Critical Critical
  • Resolution: Unresolved
  • Affects Version/s: Release 1.6
  • Fix Version/s: None
  • Component/s: None
  • Labels:
  • Patch:
    Code and Test
  • Approval:
    Triaged

Description

Clojure's random functions currently use Math.random and related features, which makes them impossible to seed. This seems like an appropriate use of a dynamic var (compared to extra arguments), since library code that wants to behave randomly could transparently support seeding without any extra effort.

I propose (def ^:dynamic *rand* (java.util.Random.)) in clojure.core, and that rand, rand-int, rand-nth, and shuffle be updated to use *rand*.

I think semantically this will not be a breaking change.

Criterium Benchmarks

I did some benchmarking to try to get an idea of the performance implications of using a dynamic var, as well as to measure the changes to concurrent access.

The code used is at https://github.com/gfredericks/clj-1452-tests; the raw output is in a comment.

rand is slightly slower, while shuffle is insignificantly faster. Using shuffle from 8 threads is insignificantly slower, but switching to a ThreadLocalRandom manually in the patched version results in a 2.5x speedup.

Running on my 8 core Linode VM:

Benchmark Clojure Runtime mean Runtime std dev
rand 1.6.0 61.3ns 7.06ns
rand 1.6.0 + *rand* 63.7ns 1.80ns
shuffle 1.6.0 12.9µs 251ns
shuffle 1.6.0 + *rand* 12.8µs 241ns
threaded-shuffling 1.6.0 151ms 2.31ms
threaded-shuffling 1.6.0 + *rand* 152ms 8.77ms
threaded-local-shuffling 1.6.0 N/A N/A
threaded-local-shuffling 1.6.0 + *rand* 64.5ms 1.41ms

Approach: create a dynamic var *rand* and update rand, rand-int, rand-nth, and shuffle to use *rand*

Patch: CLJ-1452.patch

Screened by:

Activity

Hide
Gary Fredericks added a comment -

Attached CLJ-1452.patch, with the same code used in the benchmarks.

Show
Gary Fredericks added a comment - Attached CLJ-1452.patch, with the same code used in the benchmarks.
Hide
Gary Fredericks added a comment -

Should we be trying to make Clojure's random functions thread-local by default while we're mucking with this stuff? We could have a custom subclass of Random that has ThreadLocal logic in it (avoiding ThreadLocalRandom because Java 6), and make that the default value of *rand*.

Show
Gary Fredericks added a comment - Should we be trying to make Clojure's random functions thread-local by default while we're mucking with this stuff? We could have a custom subclass of Random that has ThreadLocal logic in it (avoiding ThreadLocalRandom because Java 6), and make that the default value of *rand*.
Hide
Alex Miller added a comment -

I think the ThreadLocal question is interesting, not sure re answer.

It would be nice if the description summarized the results of the tests in a table and the criterium output was in the comments instead.

Show
Alex Miller added a comment - I think the ThreadLocal question is interesting, not sure re answer. It would be nice if the description summarized the results of the tests in a table and the criterium output was in the comments instead.
Hide
Gary Fredericks added a comment -

Full output from the test repo (which is summarized in the table in the description):

$ echo "Clojure 1.6.0"; lein with-profile +clj-1.6 run; echo "Clojure 1.6.0 with *rand*"; lein with-profile +clj-1452 run
Clojure 1.6.0

;;;;;;;;;;;;;;;;;;
;; Testing rand ;;
;;;;;;;;;;;;;;;;;;
WARNING: Final GC required 1.261632096547911 % of runtime
Evaluation count : 644646900 in 60 samples of 10744115 calls.
             Execution time mean : 61.297605 ns
    Execution time std-deviation : 7.057249 ns
   Execution time lower quantile : 56.872437 ns ( 2.5%)
   Execution time upper quantile : 84.483045 ns (97.5%)
                   Overhead used : 16.319772 ns

Found 6 outliers in 60 samples (10.0000 %)
    low-severe   1 (1.6667 %)
    low-mild     5 (8.3333 %)
 Variance from outliers : 75.5119 % Variance is severely inflated by outliers

;;;;;;;;;;;;;;;;;;;;;
;; Testing shuffle ;;
;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 4780800 in 60 samples of 79680 calls.
             Execution time mean : 12.873832 µs
    Execution time std-deviation : 251.388257 ns
   Execution time lower quantile : 12.526871 µs ( 2.5%)
   Execution time upper quantile : 13.417559 µs (97.5%)
                   Overhead used : 16.319772 ns

Found 3 outliers in 60 samples (5.0000 %)
    low-severe   3 (5.0000 %)
 Variance from outliers : 7.8591 % Variance is slightly inflated by outliers

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Testing threaded-shuffling ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 420 in 60 samples of 7 calls.
             Execution time mean : 150.863290 ms
    Execution time std-deviation : 2.313755 ms
   Execution time lower quantile : 146.621548 ms ( 2.5%)
   Execution time upper quantile : 155.218897 ms (97.5%)
                   Overhead used : 16.319772 ns
Clojure 1.6.0 with *rand*

;;;;;;;;;;;;;;;;;;
;; Testing rand ;;
;;;;;;;;;;;;;;;;;;
Evaluation count : 781707720 in 60 samples of 13028462 calls.
             Execution time mean : 63.679152 ns
    Execution time std-deviation : 1.798245 ns
   Execution time lower quantile : 61.414851 ns ( 2.5%)
   Execution time upper quantile : 67.412204 ns (97.5%)
                   Overhead used : 13.008428 ns

Found 3 outliers in 60 samples (5.0000 %)
    low-severe   3 (5.0000 %)
 Variance from outliers : 15.7596 % Variance is moderately inflated by outliers

;;;;;;;;;;;;;;;;;;;;;
;; Testing shuffle ;;
;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 4757940 in 60 samples of 79299 calls.
             Execution time mean : 12.780391 µs
    Execution time std-deviation : 240.542151 ns
   Execution time lower quantile : 12.450218 µs ( 2.5%)
   Execution time upper quantile : 13.286910 µs (97.5%)
                   Overhead used : 13.008428 ns

Found 1 outliers in 60 samples (1.6667 %)
    low-severe   1 (1.6667 %)
 Variance from outliers : 7.8228 % Variance is slightly inflated by outliers

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Testing threaded-shuffling ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 420 in 60 samples of 7 calls.
             Execution time mean : 152.471310 ms
    Execution time std-deviation : 8.769236 ms
   Execution time lower quantile : 147.954789 ms ( 2.5%)
   Execution time upper quantile : 161.277200 ms (97.5%)
                   Overhead used : 13.008428 ns

Found 3 outliers in 60 samples (5.0000 %)
    low-severe   3 (5.0000 %)
 Variance from outliers : 43.4058 % Variance is moderately inflated by outliers

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Testing threaded-local-shuffling ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 960 in 60 samples of 16 calls.
             Execution time mean : 64.462853 ms
    Execution time std-deviation : 1.407808 ms
   Execution time lower quantile : 62.353265 ms ( 2.5%)
   Execution time upper quantile : 67.197368 ms (97.5%)
                   Overhead used : 13.008428 ns

Found 1 outliers in 60 samples (1.6667 %)
    low-severe   1 (1.6667 %)
 Variance from outliers : 9.4540 % Variance is slightly inflated by outliers
Show
Gary Fredericks added a comment - Full output from the test repo (which is summarized in the table in the description):
$ echo "Clojure 1.6.0"; lein with-profile +clj-1.6 run; echo "Clojure 1.6.0 with *rand*"; lein with-profile +clj-1452 run
Clojure 1.6.0

;;;;;;;;;;;;;;;;;;
;; Testing rand ;;
;;;;;;;;;;;;;;;;;;
WARNING: Final GC required 1.261632096547911 % of runtime
Evaluation count : 644646900 in 60 samples of 10744115 calls.
             Execution time mean : 61.297605 ns
    Execution time std-deviation : 7.057249 ns
   Execution time lower quantile : 56.872437 ns ( 2.5%)
   Execution time upper quantile : 84.483045 ns (97.5%)
                   Overhead used : 16.319772 ns

Found 6 outliers in 60 samples (10.0000 %)
    low-severe   1 (1.6667 %)
    low-mild     5 (8.3333 %)
 Variance from outliers : 75.5119 % Variance is severely inflated by outliers

;;;;;;;;;;;;;;;;;;;;;
;; Testing shuffle ;;
;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 4780800 in 60 samples of 79680 calls.
             Execution time mean : 12.873832 µs
    Execution time std-deviation : 251.388257 ns
   Execution time lower quantile : 12.526871 µs ( 2.5%)
   Execution time upper quantile : 13.417559 µs (97.5%)
                   Overhead used : 16.319772 ns

Found 3 outliers in 60 samples (5.0000 %)
    low-severe   3 (5.0000 %)
 Variance from outliers : 7.8591 % Variance is slightly inflated by outliers

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Testing threaded-shuffling ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 420 in 60 samples of 7 calls.
             Execution time mean : 150.863290 ms
    Execution time std-deviation : 2.313755 ms
   Execution time lower quantile : 146.621548 ms ( 2.5%)
   Execution time upper quantile : 155.218897 ms (97.5%)
                   Overhead used : 16.319772 ns
Clojure 1.6.0 with *rand*

;;;;;;;;;;;;;;;;;;
;; Testing rand ;;
;;;;;;;;;;;;;;;;;;
Evaluation count : 781707720 in 60 samples of 13028462 calls.
             Execution time mean : 63.679152 ns
    Execution time std-deviation : 1.798245 ns
   Execution time lower quantile : 61.414851 ns ( 2.5%)
   Execution time upper quantile : 67.412204 ns (97.5%)
                   Overhead used : 13.008428 ns

Found 3 outliers in 60 samples (5.0000 %)
    low-severe   3 (5.0000 %)
 Variance from outliers : 15.7596 % Variance is moderately inflated by outliers

;;;;;;;;;;;;;;;;;;;;;
;; Testing shuffle ;;
;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 4757940 in 60 samples of 79299 calls.
             Execution time mean : 12.780391 µs
    Execution time std-deviation : 240.542151 ns
   Execution time lower quantile : 12.450218 µs ( 2.5%)
   Execution time upper quantile : 13.286910 µs (97.5%)
                   Overhead used : 13.008428 ns

Found 1 outliers in 60 samples (1.6667 %)
    low-severe   1 (1.6667 %)
 Variance from outliers : 7.8228 % Variance is slightly inflated by outliers

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Testing threaded-shuffling ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 420 in 60 samples of 7 calls.
             Execution time mean : 152.471310 ms
    Execution time std-deviation : 8.769236 ms
   Execution time lower quantile : 147.954789 ms ( 2.5%)
   Execution time upper quantile : 161.277200 ms (97.5%)
                   Overhead used : 13.008428 ns

Found 3 outliers in 60 samples (5.0000 %)
    low-severe   3 (5.0000 %)
 Variance from outliers : 43.4058 % Variance is moderately inflated by outliers

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Testing threaded-local-shuffling ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
Evaluation count : 960 in 60 samples of 16 calls.
             Execution time mean : 64.462853 ms
    Execution time std-deviation : 1.407808 ms
   Execution time lower quantile : 62.353265 ms ( 2.5%)
   Execution time upper quantile : 67.197368 ms (97.5%)
                   Overhead used : 13.008428 ns

Found 1 outliers in 60 samples (1.6667 %)
    low-severe   1 (1.6667 %)
 Variance from outliers : 9.4540 % Variance is slightly inflated by outliers
Hide
Gary Fredericks added a comment -

I think using a ThreadLocal is logically independent from adding *rand*, so it could be a separate ticket. I just suggested it here since it would for some uses mitigate any slowdown from *rand* but now that I'm looking at the benchmark results again the slowdown might be insignificant.

Show
Gary Fredericks added a comment - I think using a ThreadLocal is logically independent from adding *rand*, so it could be a separate ticket. I just suggested it here since it would for some uses mitigate any slowdown from *rand* but now that I'm looking at the benchmark results again the slowdown might be insignificant.
Hide
Gary Fredericks added a comment -

Also worth noting that (as I did in the benchmark code) with just the patch's changes (i.e., no ThreadLocal involved) users still gain the ability to do ThreadLocal manually, which is not currently possible.

Show
Gary Fredericks added a comment - Also worth noting that (as I did in the benchmark code) with just the patch's changes (i.e., no ThreadLocal involved) users still gain the ability to do ThreadLocal manually, which is not currently possible.

People

Vote (6)
Watch (2)

Dates

  • Created:
    Updated: