Open sbinet opened 2 years ago
I was just about to post a performance question (apparently not the same file, but the branch is exactly the same type):
julia> using UnROOT
[ Info: Precompiling UnROOT [3cd96dde-e98d-4713-81e9-a4a1b0235ce9]
julia> const txrd = LazyTree("root://eospublic.cern.ch//eos/root-eos/cms_opendata_2012_nanoaod/Run2012B_DoubleMuParked.root", "Events");
julia> @time sum(txrd.nMuon[1:10^5])
1.723536 seconds (3.45 M allocations: 173.713 MiB, 2.67% gc time, 29.63% compilation time)
0x0000000000039bbb
julia> @time sum(txrd.nMuon[1:10^5])
1.174588 seconds (1.27 k allocations: 1.049 MiB)
0x0000000000039bbb
julia> const thttp = LazyTree("https://jiling.web.cern.ch/jiling/public/Run2012B_DoubleMuParked.root", "Events");
julia> @time sum(thttp.nMuon[1:10^5])
0.847792 seconds (146.22 k allocations: 9.412 MiB, 8.69% compilation time)
0x0000000000039647
julia> @time sum(thttp.nMuon[1:10^5])
0.287621 seconds (752 allocations: 1.516 MiB)
0x0000000000039647
julia> @time sum(thttp.nMuon[1:10^5])
0.279659 seconds (719 allocations: 1.516 MiB)
0x0000000000039647
both of these have the same underlying logic beyond I/O source/sink. The HTTP one is not using Multipart
GET so it's a fair comparison. Both of these are "CERN" but evidently not the same network, before I've noticed eospublic
have pretty good infrastructure for the root://
so it's telling that it's this much slow.
One thing is very different is root://
doesn't seem to benefit from async
call, I wonder if it's because I'm not using Go library correctly: should ReadAt()
be capable of being called in parallel?
$> time root-dump root://ccxrootdgotest.in2p3.fr:9001/tmp/rootio/testdata/SMHiggsToZZTo4L.root > /dev/null
real 0m7.279s
user 0m8.221s
sys 0m1.256s
$> time root-dump https://cern.ch/binet/big-file.root > /dev/null
real 0m5.454s
user 0m6.156s
sys 0m0.228s
(and that's with the https
plugin downloading the whole file locally and serving it as a local file)
so there are still some performance bits to recoup. but the 10x drop has been fixed.
this is possibly related to #399.
consider:
running over the same file but with
http[s]://
(which, currently, downloads the whole file and then serves it locally):a factor 10x is left on the floor. (ok, not the same machine, on different networks, etc... but still)