HepJet

A library to benchmark various jet algorithms for HEP physics

All algorithms use double precision and rapidity-phi space to define distances. The directories to perform benchmarks are organized as following:

<p> <b>fastjet</b> - the official FastJet implementation <br> <b>ktjet</b> - Original KtJet (C++) benchmark code. Works in the anti-KT mode<br> <b>nlojet</b> - implementation of kT-jets from NLOjet++ <br> <b>jetn2_java</b> - Java implementation of the jet algorithm by I.Pogrebnyak using the N*N FastJet approach. <br> <b>scjet_cpp</b> - An alternative (traditional) implementation of kT-jet / anti-kT (N*N*N)<br> <b>scjet_java</b> - implementation of the SCJet jet algorithm in Java. <br> <b>benchmark</b> - compare FastJet and SCJet implementations using the same input.<br> <p> All algorithms run the anti-kT jet algorithm. One can also run the standard kT and Cambridge/Aachen jet algorithms. <p>

FastJet and KtJet C++ codes used by this library are taken from the orinal <a href="http://fastjet.fr/">FastJet</a> and <a href="https://ktjet.hepforge.org/">KtJet</a> web pages. JetN2 is a light-weight implementation of the anti-kT algorithms for jet validation used by the <a href="http://atlaswww.hep.anl.gov/hepsim/">HepSim</a> Monte Carlo database. More details can be found in <a href="https://github.com/chekanov/hephysics">HePhysics package</a>. The algorithm was implemented by I.Pogrebnyak in https://github.com/ivankp/javaN2jet. The Java implementation of SCJet and JetN2 is available from the <a href="http://jwork.org/scavis/">SCaVis data-analysis</a> program.

<h2>Benchmark results</h2>

The benchmarks have been done on Xeon(R) CPU E5520 @ 2.27GHz using the input file with 1000 particles (data/single-event.dat) for a 100 pp collision event (ttbar, pT>8 TeV). The processing time:

<ul> <li>fastjet - 3 msec </li> <li>jetn2_java - 15 msec using FastJet N** approach</li> <li>scjet_cpp - 105 msec (fast N^2 mode) or 990 msec (the standard N^3 mode) </li> <li>scjet_java - 103 msec (after multiple runs, fast mode) </li> <li>nlojet - ? (failed). Used by NLOjet++</li> <li>ktjet - 370 msec</li> <li>pgjet_cpp - 4756 ms msec (no cache, using std::set)</li> </ul>

In summary: <b>fastjet</b> is about a factor 30 faster than <b>scjet_cpp</b> (seeded). The Java implementation (<b>scjet_java</b>) is as fast as the C++ version when using more than one run
over events (first run is a factor 4 slower than for the C++ version due to the JIT compilation). Other similar algorithms are slower. When using the kT mode, the fast mode is disabled and <b>scjet_java</b> has the same speed as <b>ktjet</b>.

<p> </p>

There are some differences in the output jets between different implementations. There is no difference between <b>scjet_cpp</b>, and <b>fastjet</b> implementations. For the seeded option,
differences can be at the level of 2% for transverse momentum of very soft jets. The difference between <b>nlojet</b> and <b>fastjet</b> is also at the level of a few percents. No difference is found between <b>fastjet</b> and <b>ktjet</b> implementations.

<p>

Read "README" files in each directory to see how to run each benchmark. Makefile files are included inside each directory.

Authors:

S.Chekanov (ANL) I.Pogrebnyak (MSU/ANL)

Feb. 2015