Some benchmarks for serializers which can serialize arbitrary Julia structures. Currently just BSON and BSONqs.
For now the benchmark is biased towards highly nested composite data types (structs).
To run the benchmark yourself. Clone the repo then cd into its
directory and do
# julia --project
julia> using serbench
julia> do_bench()
The save benchmark for the original BSON library is disabled because it didn't complete after a few minutes most likely due to this. Performance should be very similar if that is merged. The largest differences are in reading.
In summary BSONqs.load is about 3.5x faster and uses 4.4x less memory than
BSON.load after the second run. During the first run it is only about 2x
faster, most likely because BSONqs makes heavy use of @generated methods.
julia> do_bench()
[ Info: Bench Save BSONqs
┌ Info: BSONqs.bson(io, foos)
│ elapsed = 9.270388092
│ allocatedMb = 2735.0
└ gctime = 3.046428065
[ Info: Bench load_compat BSONqs Document
┌ Info: BSONqs.load_compat(io)
│ elapsed = 2.610610094
│ allocatedMb = 499.0
└ gctime = 0.632833662
[ Info: Bench load BSONqs Document
┌ Info: BSONqs.load(io)
│ elapsed = 2.019496837
│ allocatedMb = 409.0
└ gctime = 0.204799674
[ Info: Bench load BSON Document
┌ Info: BSON.load(io)
│ elapsed = 3.926433859
│ allocatedMb = 1514.0
└ gctime = 1.247946657
julia> do_bench()
[ Info: Bench Save BSONqs
┌ Info: BSONqs.bson(io, foos)
│ elapsed = 7.343643741
│ speedup = 1.2623689845196153
│ allocatedMb = 2525.0
└ gctime = 2.728705096
[ Info: Bench load_compat BSONqs Document
┌ Info: BSONqs.load_compat(io)
│ elapsed = 2.013142543
│ speedup = 1.2967835303453723
│ allocatedMb = 419.0
└ gctime = 0.63647315
[ Info: Bench load BSONqs Document
┌ Info: BSONqs.load(io)
│ elapsed = 0.921116403
│ speedup = 2.192444766397239
│ allocatedMb = 328.0
└ gctime = 0.173687222
[ Info: Bench load BSON Document
┌ Info: BSON.load(io)
│ elapsed = 3.474502083
│ speedup = 1.1300709469167411
│ allocatedMb = 1454.0
└ gctime = 1.226204899