Fix compression benchmarks

[kamahori]

Fixed some problems about LHCb/NanoAOD/ATLAS benchmarks.

To compress properly,

• Make them use hadd command.
• Make them delete files after benchmarking not to suppress storage capacity.

Also, to save time,

• Change number of iterations from 5 to 3.
• Change file used in NanoAOD BM to smaller one (Run2012B_DoubleElectron.root).
• Change file used in ATLAS BM to 2 smaller files: atlasopendata_mc_117050.ttbar_lep.root (Monte Carlo data) and atlasopendata_DataMuons.root (real data).

[kamahori]

Output of these benchmarks can be found here.

[vgvassilev]

We should not use the disk/ssd in cases where we do not want to measure their effects. The benefits are mostly two:

• Using ramdisk is usually faster and thus we can execute more benchmarks for given time;
• We increase the life of the benchmarking machines -- usually they are expensive, physical hard-to-get machines.

Could you use GetTempFs and operate on the ramdisk?

[oshadura]

@stwunsch do you think data sets (e,g, ATLAS) and timing of benchmark is appropriate?

[stwunsch]

Could you tell me how long the benchmarks run?

[oshadura]

@stwunsch here is output from Keisuke: #179 (comment)

[oshadura]

@kamahori is it really take 122 min for Atlas benchmark? @stwunsch I will better checkout PR by myself 👍

[kamahori]

@oshadura Yes it actually took about 2hrs with my machine.

[stwunsch]

i thought we target a runtime of about 30s per benchmark ;) is this realistic for these kind of benchmarks?

[stwunsch]

I think you can set the repetitions down to 1. We get the statistcs anyway by running twice a day!

[stwunsch]

Then, here is a small NanoAOD(-like) sample: https://root.cern/files/rootbench/SMHiggsToZZTo4L.root

[kamahori]

How about number of levels? Do we really need to measure with 3 different levels?

[guitargeek]

I rebased the PR on master so we can merge it. Thank you for the improvements!