Deserialization is too slow (for the wrong reasons)

[fieker]

Try:

   save("/tmp/bla", identity_matrix(ZZ, 3000))
   load("/tmp/bla");

this takes (the load) > 30 sec. If you profile the load and look at the c-functions called, then >90% are spent in julia trying to figure out types. Less than 5% are json/ flint.

[lgoettgens]

I had a quick look and it seems that load_object does have the intended return type as second argument, but in many cases cannot infer it itself (and thus has to figure out types in followup code with a huge performance loss). A simple patch like the following already speeds things up by 25% for me:

diff --git a/src/Serialization/main.jl b/src/Serialization/main.jl
index fff1110bf73..ffd10fee676 100644
--- a/src/Serialization/main.jl
+++ b/src/Serialization/main.jl
@@ -242,15 +242,15 @@ function load_typed_object(s::DeserializerState; override_params::Any = nothing)
   end
 end
 
-function load_object(s::DeserializerState, T::Type, key::Union{Symbol, Int})
+function load_object(s::DeserializerState, T::Type{innerT}, key::Union{Symbol, Int}) where innerT
   load_node(s, key) do _
-    load_object(s, T)
+    load_object(s, T)::innerT
   end
 end
 
-function load_object(s::DeserializerState, T::Type, params::Any, key::Union{Symbol, Int})
+function load_object(s::DeserializerState, T::Type{innerT}, params::Any, key::Union{Symbol, Int}) where innerT
   load_node(s, key) do _
-    load_object(s, T, params)
+    load_object(s, T, params)::innerT
   end
 end

But since there is a major overhaul in the works by @antonydellavecchia right now (see #4162), I would first wait for that and how it impacts performance, and then we can benchmark things like my patch again and include maybe something like this if it helps.

[antonydellavecchia]

I had a quick look and it seems that load_object does have the intended return type as second argument, but in many cases cannot infer it itself (and thus has to figure out types in followup code with a huge performance loss). A simple patch like the following already speeds things up by 25% for me:

diff --git a/src/Serialization/main.jl b/src/Serialization/main.jl
index fff1110bf73..ffd10fee676 100644
--- a/src/Serialization/main.jl
+++ b/src/Serialization/main.jl
@@ -242,15 +242,15 @@ function load_typed_object(s::DeserializerState; override_params::Any = nothing)
end
end

-function load_object(s::DeserializerState, T::Type, key::Union{Symbol, Int})
+function load_object(s::DeserializerState, T::Type{innerT}, key::Union{Symbol, Int}) where innerT
load_node(s, key) do _

• load_object(s, T)

• load_object(s, T)::innerT
  end
  end

-function load_object(s::DeserializerState, T::Type, params::Any, key::Union{Symbol, Int})
+function load_object(s::DeserializerState, T::Type{innerT}, params::Any, key::Union{Symbol, Int}) where innerT
load_node(s, key) do _

• load_object(s, T, params)

• load_object(s, T, params)::innerT
  end
  end
  But since there is a major overhaul in the works by @antonydellavecchia right now (see #4162), I would first wait for that and how it impacts performance, and then we can benchmark things like my patch again and include maybe something like this if it helps.

#4162 does such a fix.
Once this is done, we can work on removing parsing for small ints