Message: Re: INCLXX plus ABLA/PRECO Not Logged In (login)
 Next-in-Thread Next-in-Thread
 Next-in-Forum Next-in-Forum

None Re: INCLXX plus ABLA/PRECO 

Forum: Hadronic Processes
Re: Note INCLXX plus ABLA/PRECO (Alexey Solovyev)
Re: None Re: INCLXX plus ABLA/PRECO (Davide Mancusi)
Re: None Re: INCLXX plus ABLA/PRECO (Alexey Solovyev)
Re: None Re: INCLXX plus ABLA/PRECO (Davide Mancusi)
Date: 04 Mar, 2016
From: Alexey Solovyev <Alexey Solovyev>

I'm running my simulation on IHEP (ihep.su) cluster and as soon I'm not cluster administrator, it may take a while to retreive fully previous running logs (I can anyway ask qsub logs if needed), but here are things I noticed last several runs:

0. I'm trying to reproduce the same data presented in http://geant4.slac.stanford.edu/Space06/presentations/03_Wednesday/05-Sasaki_ValidationOfIonPhysicsInGeant4AgainstCarbon.pdf p. 24 for different energies and also to differentiate each isotope contribution. So I have lots root histos, DetectorSD to put data depending on step->GetTrack()->GetParticleDefinition()->GetAtomicNumber/Mass, and 1D replica of 0.5-1mm G4Tubs in my geometry. So e.g. for 1e7 primary particles in single thread in Windows 7 (as my developer machine) the highest memory consumption of this simulation is approx. 1.2 Gb.

1. I have 5Gb vmem limit (on cluster) for the whole task (no matter how much nodes I'm using). First I found that all my tasks which were using QGSP_INCLXX we killed by memory limit while code linked with 10.2.

2. While investigating this issue I also found that e.g. QGSP_BIC/QBBC/Shielding don't die by memory in 10.2.

3. For the chosen nps (1e7) using QGSP_BIC in 10.2 took about 300 hrs. cpu time (in 5 nodes, 60 hrs. wall-time) and 330 hrs. in 10.1. For the QGSP_INCLXX it tooks about 322 hrs. cpu-time in 10.1 and killed by memory limit in 10.2. When killed it completed approx. 3e6 nps, and while similar tasks runned I got with qstat -f that after 144. hrs cpu-time it utilized almost 4.3 Gb (and killed a bit later, again I can't now directly access kill logs so can't say now how much cpu/wall time passed before it has been killed and I can't say how much vmem used completed tasks in the past, but the completed tasks definitely not overcome 5Gb vmem limit). Actually all my simulations with QGSP_INCLXX were killed by vmem limit with same nps (1e3 on ui node and 1e6 worked succesfully btw) in all my geometries in 10.2 (I have single file for all cases, switching geometry depending on messenger or command line, but I think it's not essential here and all other geometries are more sophisticated).

4. I'm not fully sure that's the INCLXX bug or something because I also store all the secondary partciles using SteppingAction the same way as in Hadr03 example and if INCLXX generate lots of excited particles this storage may overcome the memory limit, but anyway this is not happening in 10.1 nor with other phys lists in 10.2.

I can ask for the run logs if needed, and maybe we better use e-mail to not flood this topic a lot.

Inline Depth:
 1 1
 All All
Outline Depth:
 1 1
 2 2
 All All
Add message: (add)

1 None: Re: INCLXX plus ABLA/PRECO   (Davide Mancusi - 04 Mar, 2016)
 Add Message Add Message
to: "Re: INCLXX plus ABLA/PRECO"

 Subscribe Subscribe

This site runs SLAC HyperNews version 1.11-slac-98, derived from the original HyperNews


[ Geant 4 Home | Geant 4 HyperNews | Search | Request New Forum | Feedback ]