Strange speeds

ruby-talk
1

Hello ruby people,
  I've tried to make some speed test in ruby to find out how fast is searching in hash (using has_key?)

The results were OK but I discovered curious thing -- when i generate big hash the speed is different when i unserialize (marshal.load) the same hash.

Unserialized Hash is 1.5x faster in has_key? method. Tried it with GC.disable as well, results were same.

What am I missing?

Thanx for any ideas

2

The first reason that comes to mind is different insertion order. But I hardly believe it could make such a big difference. Care to post your testing code?

Kind regards

    robert

···

kryglik <kryglik@iol.cz> wrote:

Hello ruby people,
I've tried to make some speed test in ruby to find out how fast is
searching in hash (using has_key?)

The results were OK but I discovered curious thing -- when i generate
big hash the speed is different when i unserialize (marshal.load) the
same hash.

Unserialized Hash is 1.5x faster in has_key? method. Tried it with
GC.disable as well, results were same.

What am I missing?

Thanx for any ideas

3

Robert Klemme wrote:

Hello ruby people,
I've tried to make some speed test in ruby to find out how fast is
searching in hash (using has_key?)

The results were OK but I discovered curious thing -- when i generate
big hash the speed is different when i unserialize (marshal.load) the
same hash.

Unserialized Hash is 1.5x faster in has_key? method. Tried it with
GC.disable as well, results were same.

What am I missing?

Thanx for any ideas

The first reason that comes to mind is different insertion order. But I hardly believe it could make such a big difference. Care to post your testing code?

Kind regards

   robert

Well, insertion order will be same, won't be?

h = {}
h_slovo =

if not true #HERE YOU SWITCH IF GENERATE/LOAD
     for x in 1..40_000
         if x % 2000 == 0
             puts x
         end

         slovo = ""

         for z in 1..32
             slovo += (rand(25)+65).chr
         end

         if rand(10) == 3
             h_slovo << slovo
         end

         h.update slovo => rand(1_000_000)
         slovo = nil
     end

     f = File.new("hash.marshal","w+")
     f.puts(Marshal.dump(h))
     f.close

     f = File.new("hash1.marshal","w+")
     f.puts(Marshal.dump(h_slovo))
     f.close
else

     puts "loading hash"
     h = Marshal.load(File.open("hash.marshal","r"))
     puts "loading hash1"
     h_slovo = Marshal.load(File.open("hash1.marshal","r"))
     puts "done"
end

t1 = Time.now

for slovo1 in h_slovo
h.has_key? h_slovo
end

t2 = Time.now

puts "hash: " + h.size.to_s
puts "hled: " + h_slovo.size.to_s
puts t2-t1

···

kryglik <kryglik@iol.cz> wrote:

4

Robert Klemme wrote:

Hello ruby people,
I've tried to make some speed test in ruby to find out how fast is
searching in hash (using has_key?)

The results were OK but I discovered curious thing -- when i
generate big hash the speed is different when i unserialize
(marshal.load) the same hash.

Unserialized Hash is 1.5x faster in has_key? method. Tried it with
GC.disable as well, results were same.

What am I missing?

Thanx for any ideas

The first reason that comes to mind is different insertion order.
But I hardly believe it could make such a big difference. Care to
post your testing code?

Kind regards

   robert

Well, insertion order will be same, won't be?

Not necessarily. If Marshal.dump uses the current order in the hash for writing then the insertion order during loading is likely to be different from the original insertion order.

h = {}
h_slovo =

if not true #HERE YOU SWITCH IF GENERATE/LOAD
    for x in 1..40_000
        if x % 2000 == 0
            puts x
        end

        slovo = ""

        for z in 1..32
            slovo += (rand(25)+65).chr
        end

        if rand(10) == 3
            h_slovo << slovo
        end

        h.update slovo => rand(1_000_000)
        slovo = nil
    end

    f = File.new("hash.marshal","w+")
    f.puts(Marshal.dump(h))
    f.close

    f = File.new("hash1.marshal","w+")
    f.puts(Marshal.dump(h_slovo))
    f.close
else

    puts "loading hash"
    h = Marshal.load(File.open("hash.marshal","r"))
    puts "loading hash1"
    h_slovo = Marshal.load(File.open("hash1.marshal","r"))
    puts "done"
end

t1 = Time.now

for slovo1 in h_slovo
h.has_key? h_slovo
end

t2 = Time.now

puts "hash: " + h.size.to_s
puts "hled: " + h_slovo.size.to_s
puts t2-t1

Your time measurement is far too unprecise on a modern system. You'll have to repeat lookups to get more accurate results. I get a difference between 5 and 10 percent with the attached code - regardless of GC enabled or disabled. Strange though that the copy is always slower. This is an interesting issue.

Note, if you freeze keys lookups in the original hash are much faster because then hash keys are not copied on insertion (an optimisation of unfrozen Strings as hash keys), i.e., during lookup if the key is the same it's also the same object which will be the first test - and of course this test is much faster than sequential comparison of characters.

Another note, a more realistic scenario would be if not 10% of all keys were used for lookup but if there also were keys that are not present in the hash.

Kind regards

    robert

hash-lookup.rb (709 Bytes)

···

kryglik <kryglik@iol.cz> wrote:

kryglik <kryglik@iol.cz> wrote:

5

Well, insertion order will be same, won't be?

This is not the same hash

moulon% cat b.rb
#!/usr/bin/ruby
h = {}
32.times {|i| h[(rand(25)+65).chr] = i}
File.open("hash.marshal","w+") {|f| f.puts(Marshal.dump(h)) }
h1 = File.open("hash.marshal","r") {|f| Marshal.load(f) }
puts h.keys.join(' ')
puts h1.keys.join(' ')
moulon%

moulon% ./b.rb
V K W L A B M X C D O E F Q G R H I U
K V A L W X M B C O D E Q F R G H I U
moulon%

Guy Decoux

6

Hello
  thanx for nice script!

I find out that i got a little error :o) [see the 'for' cycle in benchmark part:
for slovo1 in h_slovo
      h.has_key? h_slovo (SHOULD NOT BE h_slovo BUT slovo1)
end

so when I run your script i got confused about speeds like 1-3s instead of my 60s. The difference between marshaled and generated was like 35s and 60s (10000 keys in 100.000 hash).

Still the mysteria continues -- why the different speeds.

And also thanx for lecture of ruby optimization - not quite a guru (yet :-D).

Benchmark is a nice module :slight_smile:
Robert Klemme wrote:

···

kryglik <kryglik@iol.cz> wrote:

Robert Klemme wrote:

kryglik <kryglik@iol.cz> wrote:

Hello ruby people,
I've tried to make some speed test in ruby to find out how fast is
searching in hash (using has_key?)

The results were OK but I discovered curious thing -- when i
generate big hash the speed is different when i unserialize
(marshal.load) the same hash.

Unserialized Hash is 1.5x faster in has_key? method. Tried it with
GC.disable as well, results were same.

What am I missing?

Thanx for any ideas

The first reason that comes to mind is different insertion order.
But I hardly believe it could make such a big difference. Care to
post your testing code?

Kind regards

   robert

Well, insertion order will be same, won't be?

Not necessarily. If Marshal.dump uses the current order in the hash for writing then the insertion order during loading is likely to be different from the original insertion order.

h = {}
h_slovo =

if not true #HERE YOU SWITCH IF GENERATE/LOAD
    for x in 1..40_000
        if x % 2000 == 0
            puts x
        end

        slovo = ""

        for z in 1..32
            slovo += (rand(25)+65).chr
        end

        if rand(10) == 3
            h_slovo << slovo
        end

        h.update slovo => rand(1_000_000)
        slovo = nil
    end

    f = File.new("hash.marshal","w+")
    f.puts(Marshal.dump(h))
    f.close

    f = File.new("hash1.marshal","w+")
    f.puts(Marshal.dump(h_slovo))
    f.close
else

    puts "loading hash"
    h = Marshal.load(File.open("hash.marshal","r"))
    puts "loading hash1"
    h_slovo = Marshal.load(File.open("hash1.marshal","r"))
    puts "done"
end

t1 = Time.now

for slovo1 in h_slovo
h.has_key? h_slovo
end

t2 = Time.now

puts "hash: " + h.size.to_s
puts "hled: " + h_slovo.size.to_s
puts t2-t1

Your time measurement is far too unprecise on a modern system. You'll have to repeat lookups to get more accurate results. I get a difference between 5 and 10 percent with the attached code - regardless of GC enabled or disabled. Strange though that the copy is always slower. This is an interesting issue.

Note, if you freeze keys lookups in the original hash are much faster because then hash keys are not copied on insertion (an optimisation of unfrozen Strings as hash keys), i.e., during lookup if the key is the same it's also the same object which will be the first test - and of course this test is much faster than sequential comparison of characters.

Another note, a more realistic scenario would be if not 10% of all keys were used for lookup but if there also were keys that are not present in the hash.

Kind regards

   robert

7

Hello
thanx for nice script!

I find out that i got a little error :o) [see the 'for' cycle in
benchmark part:
for slovo1 in h_slovo
     h.has_key? h_slovo (SHOULD NOT BE h_slovo BUT slovo1)
end

so when I run your script i got confused about speeds like 1-3s
instead of my 60s. The difference between marshaled and generated was
like 35s and 60s (10000 keys in 100.000 hash).

In this case it's probably due to hash calculation for the hash you accidentally used as key.

Still the mysteria continues -- why the different speeds.

Yeah, I guess we would have to dive into the sources to find out...

And also thanx for lecture of ruby optimization - not quite a guru
(yet :-D).

Take your time. :slight_smile:

Benchmark is a nice module :slight_smile:

It really is!

Kind regards

    robert

···

kryglik <kryglik@gmail.com> wrote: