make branch basket buffer threadlocal

[aminnj]

In each of the branch objects, the last-read basket data is stored. If one tried to naively @threads an event loop, there would be huge contention in this buffer cache. This PR makes nthreads() such buffer caches.

using UnROOT
using LVCyl
using FHist
using Base.Threads

const f = ROOTFile("Run2012BC_DoubleMuParked_Muons.root")
const t = LazyTree(f, "Events", ["nMuon", r"^Muon_(pt|eta|phi|mass|charge)"])

const hmass = Hist1D(Int64; bins=0:0.001:150)
@time @threads for (i,evt) in enumerate(t)
    evt.nMuon != 2 && continue
    sum(evt.Muon_charge) != 0 && continue
    lvs = LorentzVectorCyl.(evt.Muon_pt, evt.Muon_eta, evt.Muon_phi, evt.Muon_mass)
    mass = sum(lvs).mass
    push!(hmass, mass) # rename to atomic_push eventually
end

gives

27.897073 seconds (49.06 M allocations: 39.579 GiB, 23.23% gc time) # with @threads
68.897655 seconds (48.65 M allocations: 38.559 GiB, 7.48% gc time) # without @threads

for julia with 4 threads on a mac laptop. That's about a 2.5x speedup (consistent with me seeing 300% cpu usage instead of 400%).

Note that I commented out the LRU cache around readbasketseek() in both cases to keep RAM usage low (seems to go beyond the advertised maxsize=1024^3 🤷)

The example comes from #73 .

[codecov]

Codecov Report

Merging #76 (08380f8) into master (ac38c75) will increase coverage by 0.54%.
The diff coverage is 75.00%.

@@            Coverage Diff             @@
##           master      #76      +/-   ##
==========================================
+ Coverage   82.55%   83.10%   +0.54%     
==========================================
  Files          10       10              
  Lines        1244     1302      +58     
==========================================
+ Hits         1027     1082      +55     
- Misses        217      220       +3     

Impacted Files	Coverage Δ
src/iteration.jl	68.46% <75.00%> (+2.35%)	⬆️
src/streamers.jl	90.31% <0.00%> (-0.05%)	⬇️
src/utils.jl	100.00% <0.00%> (ø)
src/custom.jl	48.86% <0.00%> (+0.02%)	⬆️
src/types.jl	92.50% <0.00%> (+0.09%)	⬆️
src/io.jl	93.93% <0.00%> (+0.18%)	⬆️
src/root.jl	90.74% <0.00%> (+0.21%)	⬆️
src/bootstrap.jl	83.56% <0.00%> (+0.23%)	⬆️
src/displays.jl	50.00% <0.00%> (+1.42%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ac38c75...08380f8. Read the comment docs.

[tamasgal]

Awesome! Btw. the memory estimation is very basic and does not catch nested data well, that's why it can grow larger than the set 1GB.

[aminnj]

Since @threads relies on getindex() to do the chunking, this previously wouldn't work:

@threads for (i,evt) in enumerate(t)
    i>5 && break
    evt.nMuon != 2 && continue
    sum(evt.Muon_charge) != 0 && continue
    # ...
end

because enumerate(t) doesn't assume a getindex() interface by default (see here), which is pretty inconvenient/unintuitive if you want to take some test code and slap on @threads. One would have to write

@threads for i in 1:length(t)
    i>5 && break
    t.nMuon[i] != 2 && continue
    sum(t.Muon_charge[i]) != 0 && continue
    # ...
end

However, our LazyTable in indexable, so based on the code here, I added some extra definitions to allow us to @threads ... enumerate(t).

[aminnj]

As far as features, I think I'm done in this PR. A review is welcome :)

[tamasgal]

Weeeha! 😃