为什么我的Trie查找比标准F＃Map的查找速度慢？

时间：2012-07-13 13:45:44

标签： performance dictionary f# inline trie

所以，我刚从OCaml移植了Trie。不幸的是，就tryFind而言，它比标准Map运行得慢。我不明白这一点 - 特里似乎应该更快。 F＃的代码库是以某种特殊的方式构建的，以使它们比用户通常部署的代码更快吗？

这是代码 - 

[<RequireQualifiedAccess>]
module Trie

type Node<'k, 'v when 'k : comparison> =
    { TrieMap : Map<'k, Node<'k, 'v>>
      TrieKvp : ('k list * 'v) option }
    member inline x.IsEmpty = x.TrieKvp.IsNone && x.TrieMap.IsEmpty

let inline make map kvp =
    { TrieMap = map
      TrieKvp = kvp }

let inline makeEmpty () : Node<'k, 'v> = make Map.empty None

let inline isEmpty (node : Node<'k, 'v>) = node.IsEmpty

let rec tryFind (key : 'k list) node =
    if key.IsEmpty then
        match node.TrieKvp with
        | Some (_, value) -> Some value
        | None -> None
    else
        let keyHead = key.Head
        let keyTail = key.Tail
        let optSubNode = Map.tryFind keyHead node.TrieMap
        match optSubNode with
        | Some subNode -> tryFind keyTail subNode
        | None -> None

let inline containsKey key node =
    (tryFind key node).IsSome

let rec addInternal (key : 'k list) value node =
    if key.IsEmpty then make node.TrieMap (Some (key, value))
    else
        let keyHead = key.Head
        let keyTail = key.Tail
        let newTrie =
            match Map.tryFind keyHead node.TrieMap with
            | Some subTrie -> subTrie
            | None -> makeEmpty ()
        let newTrie2 = addInternal keyTail value newTrie
        make (Map.add keyHead newTrie2 node.TrieMap) node.TrieKvp

let inline add key value node =
    addInternal key value node

let rec addMany kvps node =
    if Seq.isEmpty kvps then node
    else
        let kvpHead = Seq.head kvps
        let kvpTail = Seq.skip 1 kvps
        let newTrie = add (fst kvpHead) (snd kvpHead) node
        addMany kvpTail newTrie

let inline ofList kvps =
    addMany kvps (makeEmpty ())

let inline ofListBy by kvps =
    let pairs = List.map by kvps
    ofList pairs

let rec foldInternal folder rev node state =
    match node.TrieKvp with
    | Some (_, value) -> folder (Map.fold (fun state key value -> foldInternal folder (key :: rev) value state) state node.TrieMap) (List.rev rev) value
    | None -> Map.fold (fun state key value -> foldInternal folder (key :: rev) value state) state node.TrieMap

let inline fold folder state node =
    foldInternal folder [] node state

let rec map (mapper : 'k list -> 'v -> 'a) (node : Node<'k, 'v>) : Node<'k, 'a> =
    match node.TrieKvp with
    | Some (key, value) -> make (Map.map (fun _ value -> map mapper value) node.TrieMap) (Some (key, mapper key value))
    | None -> make (Map.map (fun _ value -> map mapper value) node.TrieMap) None

let inline toValueList node =
    fold (fun state _ value -> value :: state) [] node

let inline singleton (key, value) =
    add key value (makeEmpty ())

这是Jon Harrop提供的性能测试，我发现它足以衡量改进 - 

let xs = Array.init 1000000 (fun i -> [i])
let timer = System.Diagnostics.Stopwatch.StartNew()
let mutable t = Trie.makeEmpty()
for i=0 to xs.Length-1 do
    t <- Trie.add xs.[i] xs.[i] t
printfn "Trie took %fs to build" timer.Elapsed.TotalSeconds
timer.Restart()
for _ in 1..100 do
    for i=0 to xs.Length-1 do
        ignore(Trie.tryFind xs.[i])
printfn "Trie took %fs to search" timer.Elapsed.TotalSeconds

let timer = System.Diagnostics.Stopwatch.StartNew()
let mutable t = Map.empty
for i=0 to xs.Length-1 do
    t <- Map.add xs.[i] xs.[i] t
printfn "Map took %fs to build" timer.Elapsed.TotalSeconds
timer.Restart()
for _ in 1..100 do
    for i=0 to xs.Length-1 do
        ignore(Map.tryFind xs.[i])
printfn "Map took %fs to search" timer.Elapsed.TotalSeconds

注意：如果您考虑到更快的查找数据结构，请注意我需要持久的数据结构。

3 个答案:

答案 0 :(得分：4)

  

不幸的是，就tryFind而言，它的运行速度比标准Map慢。我不明白这一点 - 特里似乎应该更快。

这里的快速基准测试表明，至少在简单的情况下，你的特里已经快于Map：

do
    let n = 0
    let xs = Array.init 1000000 (fun i -> [i])
    let timer = System.Diagnostics.Stopwatch.StartNew()
    let mutable t = Trie.makeEmpty()
    for i=0 to xs.Length-1 do
        t <- Trie.add xs.[i] xs.[i] t
    printfn "Trie took %fs to build" timer.Elapsed.TotalSeconds
    timer.Restart()
    for _ in 1..100 do
        for i=0 to xs.Length-1 do
            ignore(Trie.tryFind xs.[i])
    printfn "Trie took %fs to search" timer.Elapsed.TotalSeconds

    let timer = System.Diagnostics.Stopwatch.StartNew()
    let mutable t = Map.empty
    for i=0 to xs.Length-1 do
        t <- Map.add xs.[i] xs.[i] t
    printfn "Map took %fs to build" timer.Elapsed.TotalSeconds
    timer.Restart()
    for _ in 1..100 do
        for i=0 to xs.Length-1 do
            ignore(Map.tryFind xs.[i])
    printfn "Map took %fs to search" timer.Elapsed.TotalSeconds

我得到4s来构建你的Trie，8.7s来构建Map和0.7来搜索这两种情况。

但是，您的实施还有很大的改进空间。我最近写了一篇关于F＃中优化的通用持久哈希trie实现的文章，该文章已发布here。

您之后的评论意味着您只想使用它来映射字符串。如果是这样的话，将字符串密钥专门化为特征会更有效率。

修改

KVB建议我详细说明“改进的空间”，所以这里有一些反馈：

  • 谨慎使用inline作为优化，并且仅基于引人注目的性能测量。
  • 使empty成为值而不是函数。
  • 尽可能避免使用List.head和List.tail。改为使用模式匹配。
  • 尽可能避免通用性（例如，在这种情况下为字符串键的硬编码）。

答案 1 :(得分：4)

好吧，经过多一点思考之后，我假设性能的真正区别在于使用键列表而不是字符串。字符串（和数组）具有更好的缓存一致性。所以，我将密钥从'k列表更改为字符串瞧！现在，性能实际上比我的应用程序中的Map更好！

这是代码 - 

[<RequireQualifiedAccess>]
module StringTrie

type Node<'v> =
    { TrieMap : Map<char, Node<'v>>
      TrieKvp : (string * 'v) option }
    member inline x.IsEmpty = x.TrieKvp.IsNone && x.TrieMap.IsEmpty

let inline make map kvp =
    { TrieMap = map
      TrieKvp = kvp }

let inline makeEmpty () : Node<'v> = make Map.empty None

let inline isEmpty (node : Node<'v>) = node.IsEmpty

let rec tryFindInternal (key : string) index node =
    if key.Length = index then
        match node.TrieKvp with
        | Some (_, value) -> Some value
        | None -> None
    else
        let optSubNode = Map.tryFind key.[index] node.TrieMap
        match optSubNode with
        | Some subNode -> tryFindInternal key (index + 1) subNode
        | None -> None

let inline tryFind (key : string) node =
    tryFindInternal key 0 node

let inline containsKey key node =
    (tryFind key node).IsSome

let rec addInternal (key : string) index value node =
    if key.Length = index then make node.TrieMap (Some (key, value))
    else
        let char = key.[index]
        let newTrie =
            match Map.tryFind char node.TrieMap with
            | Some subTrie -> subTrie
            | None -> makeEmpty ()
        let newTrie2 = addInternal key (index + 1) value newTrie
        make (Map.add char newTrie2 node.TrieMap) node.TrieKvp

let inline add key value node =
    addInternal key 0 value node

let rec addMany kvps node =
    if Seq.isEmpty kvps then node
    else
        let kvpHead = Seq.head kvps
        let kvpTail = Seq.skip 1 kvps
        let newTrie = add (fst kvpHead) (snd kvpHead) node
        addMany kvpTail newTrie

let inline ofList kvps =
    addMany kvps (makeEmpty ())

let inline ofListBy by kvps =
    let pairs = List.map by kvps
    ofList pairs

let rec foldInternal folder rev node state =
    match node.TrieKvp with
    | Some (_, value) -> folder (Map.fold (fun state key value -> foldInternal folder (key :: rev) value state) state node.TrieMap) (List.rev rev) value
    | None -> Map.fold (fun state key value -> foldInternal folder (key :: rev) value state) state node.TrieMap

let inline fold folder state node =
    foldInternal folder [] node state

let rec map (mapper : string -> 'v -> 'a) (node : Node<'v>) : Node<'a> =
    match node.TrieKvp with
    | Some (key, value) -> make (Map.map (fun _ value -> map mapper value) node.TrieMap) (Some (key, mapper key value))
    | None -> make (Map.map (fun _ value -> map mapper value) node.TrieMap) None

let inline toValueList node =
    fold (fun state _ value -> value :: state) [] node

let inline singleton (key, value) =
    add key value (makeEmpty ())

我还构建了一个适用于数组的版本，并且速度很快 - 

[<RequireQualifiedAccess>]
module ArrayTrie

type Node<'k, 'v when 'k : comparison> =
    { TrieMap : Map<'k, Node<'k, 'v>>
      TrieKvp : ('k array * 'v) option }
    member inline x.IsEmpty = x.TrieKvp.IsNone && x.TrieMap.IsEmpty

let inline make map kvp =
    { TrieMap = map
      TrieKvp = kvp }

let inline makeEmpty () : Node<'k, 'v> = make Map.empty None

let inline isEmpty (node : Node<'k, 'v>) = node.IsEmpty

let rec tryFindInternal (key : 'k array) index node =
    if key.Length = index then
        match node.TrieKvp with
        | Some (_, value) -> Some value
        | None -> None
    else
        let optSubNode = Map.tryFind key.[index] node.TrieMap
        match optSubNode with
        | Some subNode -> tryFindInternal key (index + 1) subNode
        | None -> None

let inline tryFind (key : 'k array) node =
    tryFindInternal key 0 node

let inline containsKey key node =
    (tryFind key node).IsSome

let rec addInternal (key : 'k array) index value node =
    if key.Length = index then make node.TrieMap (Some (key, value))
    else
        let char = key.[index]
        let newTrie =
            match Map.tryFind char node.TrieMap with
            | Some subTrie -> subTrie
            | None -> makeEmpty ()
        let newTrie2 = addInternal key (index + 1) value newTrie
        make (Map.add char newTrie2 node.TrieMap) node.TrieKvp

let inline add key value node =
    addInternal key 0 value node

let rec addMany kvps node =
    if Seq.isEmpty kvps then node
    else
        let kvpHead = Seq.head kvps
        let kvpTail = Seq.skip 1 kvps
        let newTrie = add (fst kvpHead) (snd kvpHead) node
        addMany kvpTail newTrie

let inline ofList kvps =
    addMany kvps (makeEmpty ())

let inline ofListBy by kvps =
    let pairs = List.map by kvps
    ofList pairs

let rec foldInternal folder rev node state =
    match node.TrieKvp with
    | Some (_, value) -> folder (Map.fold (fun state key value -> foldInternal folder (key :: rev) value state) state node.TrieMap) (List.rev rev) value
    | None -> Map.fold (fun state key value -> foldInternal folder (key :: rev) value state) state node.TrieMap

let inline fold folder state node =
    foldInternal folder [] node state

let rec map (mapper : 'k array -> 'v -> 'a) (node : Node<'k, 'v>) : Node<'k, 'a> =
    match node.TrieKvp with
    | Some (key, value) -> make (Map.map (fun _ value -> map mapper value) node.TrieMap) (Some (key, mapper key value))
    | None -> make (Map.map (fun _ value -> map mapper value) node.TrieMap) None

let inline toValueList node =
    fold (fun state _ value -> value :: state) [] node

let inline singleton (key, value) =
    add key value (makeEmpty ())

唯一能够提高性能的东西就是获得一个指向字符串的内部指针，而不是一遍又一遍地执行索引。这在F＃中似乎并不容易，但对C＃中的数组来说似乎至少是可能的。

答案 2 :(得分：2)

为什么不呢？ OCaml怎么样，那有没有更快？由于Trie是按Map实现的，因此我预计至少会有一些输入慢于Map。在某些情况下，它仍然可能优于Map，例如当尺寸非常大时。

另外，如果您的主要兴趣是查找性能，为什么不冻结您的Trie以使用基于Dictionary的节点？

相关问题
最新问题