Wow - that's a pretty big win. I think we should try and get this into
Clojure ASAP.
Are we too late for 1.6?
On Sunday, 16 February 2014 18:48:09 UTC+8, Jules wrote:
>
> Thanks, Mikera
>
> You are right about merge:
>
> user=> (def m1 (apply hash-map (range 10000000)))
> #'user/m1
> user=> (def m2 (apply hash-map (range 5000000 15000000)))
> #'user/m2
> user=> (time (def m3 (merge m1 m2)))
> "Elapsed time: 5432.184582 msecs"
> #'user/m3
> user=> (time (def m4 (clojure.lang.PersistentHashMap/splice m1 m2)))
> "Elapsed time: 1064.268269 msecs"
> #'user/m4
> user=> (= m3 m4)
> true
> user=>
>
> as you would expect, a splice is faster and causes less of a memory spike.
>
>
> Jules
>
>
> On Sunday, 16 February 2014 10:01:04 UTC, Mikera wrote:
>>
>> +1 for this approach - I've wanted something like this several times.
>>
>> It's only an "optimisation", but it's a very useful one. Same technique
>> can probably be used to accelerate "merge" significantly which is a pretty
>> common operation when you are building map-like structures.
>>
>> On Sunday, 16 February 2014 07:06:24 UTC+8, Jules wrote:
>>>
>>> Guys,
>>>
>>> I've been playing with reducers on and off for a while but have been
>>> frustrated because they don't seem to fit a particular usecase that I have
>>> in mind... specifically: getting as many associations into a hash-map as as
>>> I can in as short a time as possible.
>>>
>>> My understanding of the reason for this is that reducers practice a
>>> divide and conquer strategy. The incoming sequence is divided up. Each
>>> sub-sequence is reduced into a sub-result (possibly in parallel) and then
>>> the sub-results are combined into the the final outgoing result.
>>>
>>> Unfortunately, there does not seem to be a better way of combining two
>>> hash-maps other than to read each entry from one and create a new
>>> corresponding association in the other. This means that each recombination
>>> in the above process essentially repeats most of the work already performed
>>> in the previous reduction stage.
>>>
>>> Hash-sets are implemented via hash-maps, and simpler with which to
>>> demonstrate this problem:
>>>
>>> user=> (def a (doall (range 10000000)))
>>> #'user/a
>>> user=> (def b (doall (range 5000000 15000000)))
>>> #'user/b
>>> user=> (time (def c (into #{} a)))
>>> "Elapsed time: 6319.392669 msecs"
>>> #'user/c
>>> user=> (time (def d (into #{} b)))
>>> "Elapsed time: 5389.805233 msecs"
>>> #'user/d
>>> user=> (time (def e (into c d)))
>>> "Elapsed time: 8486.032191 msecs"
>>> #'user/e
>>>
>>>
>>> In the example above, you can see that the reduction into hash-sets of
>>> two overlapping lists of 10,000,000 elements takes 6.3 and 5.4 seconds.
>>> This stage can be carried out in parallel i.e. time elapsed for this stage
>>> would be 6.3 seconds - but we now have two hash-sets and we want one, so we
>>> have to combine them.
>>>
>>>
>>> user=> (time (def e (into c d)))
>>> "Elapsed time: 8486.032191 msecs"
>>> #'user/e
>>>
>>> As you can see, all the advantages of splitting the original sequence
>>> into 2 and processing the two halves in parallel are lost since the
>>> recombination or their results takes 8.5 seconds - more than we saved by
>>> doing the reduction in parallel.
>>>
>>> So, what can we do about it ?
>>>
>>> I had a look at the code for PersistentHashMap (PersistentHashSet uses
>>> PersistantHashMap internally). I realised that it was possible to "splice"
>>> together the internal structure of two hash maps into a single one without
>>> repeating most of the work required to build one from scratch. So, I had a
>>> go at implementing it:
>>>
>>>
>>> user=> (time (def f (clojure.lang.PersistentHashSet/splice c d)))
>>> "Elapsed time: 3052.690911 msecs"
>>> #'user/f
>>>
>>> and:
>>>
>>> user=> (= e f)
>>> true
>>>
>>> Whilst this is still adding 3 seconds to our time, that 3 seconds is
>>> half the time that we would have added had we executed the second reduction
>>> in serial, rather than in parallel.
>>>
>>> This means that we can now reduce large datasets into sets/maps more
>>> quickly in parallel than we can in serial :-) As an added benefit, because
>>> splice reuses as much of the internal structure of both inputs as possible,
>>> it's impact in terms of heap consumption and churn is less - although I
>>> think that a full implementation might add some Java-side code complexity.
>>>
>>> If you would like to give 'splice' a try out, you will need to clone my
>>> fork of clojure at github
>>>
>>> https://github.com/JulesGosnell/clojure
>>>
>>> Please let me know if you try out the code. I would be interested to
>>> hear if people think it is worth pursuing.
>>>
>>> I was also thinking that it should be possible to use a similar trick to
>>> quickly and cheaply split a map/set into [roughly] equal sized pieces
>>> (assuming an good hash distribution). This would enable the use of a
>>> map/set as an input sequence into the parallel reduction process outlined
>>> above. Currently, I believe that only a vector can be used in this way. It
>>> would be harder to arrange that 'count' could be implemented efficiently on
>>> these sub-maps/sets, but this is not important for the reduction process.
>>>
>>> BTW - benchmarks were run on a 3.2ghz Phenom II / clojure/master /
>>> openjdk-1.7.0_51 / Fedora 20 with min and max 4gb ram.
>>>
>>> regards,
>>>
>>>
>>>
>>> Jules
>>>
>>>
>>>
--
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to [email protected]
Note that posts from new members are moderated - please be patient with your
first post.
To unsubscribe from this group, send email to
[email protected]
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
---
You received this message because you are subscribed to the Google Groups
"Clojure" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/groups/opt_out.
