Minutes of the ICE section
24th meeting on Wednesday 13/04/2011 (08:40-10:30, 6/2-004)
ICE members: Benoit Salvant (BS), Christian Hansen (CH), Carlo Zannini (CZ), Hugo Alistair Day (HD), Elena Benedetto (EB), Ewen Hamish Maclean (EHM), Elias Metral (EM), Elena Wildner (EW), Frank Schmidt (FS), Giovanni Rumolo (GR), Jean-Luc Nougaret (JLN), Kevin Shing Bruce Li (KL), Maria Carmen Alabau Pons (MCAP), Nicolo Biancacci (NB), Nicolas Mounet (NM), Olav Ejner Berrig (OB), Tatiana Pieloni (TP), Werner Herr (WH).
Present/Excused: BS, CH, CZ, HD, EB, EHM, EM, EW, FS, GR, JLN, KL, MCAP, NB, NM, OB, TP, WH, Alessandro Vivoli, Chandrashekhara Bhat, Fritz Caspers, Hannes Bartosik, Harry Renshall, Massimo Giovannozzi, Rama Calaga, Tom Mertens, Vittorio Vaccaro.
1) Newcomers / visitors
- None.
2) Comments on the minutes of the previous 23th meeting + Actions
- No comment.
3) General infos
- GR mentioned some problems with the PSB ring 1 since few days (and now it seems that ring 2 is also degrading):
- See ElogBook on Monday 11/04/11: https://ab-dep-op-elogbook.web.cern.ch/ab-dep-op-elogbook/elogbook/eLogbook.php?shiftId=1034164.
- Appears already with low-intensity beams (3E12 p/p).
- Without the Transverse Feedback (TF), 3 losses can be observed: the first 2 are old friends (when the TF is not ON) but the 3rd one is new and in particular in ring 1 (in fact it is not seen in the other rings) and especially in the H-p lane.
- Is this a problem linked to a new impedance introduced (for instance during the Technical Stop) or a HW issue (linked to the TF)?
- Reminder: In the past, a 3rd instability was also sometimes observed when the TF was OFF, but it was later in the cycle and solved since then.
- Reminder: Today is the deadline for the abstract submission to IPAC2011!
- SL meeting:
- The 5% flexibility mentioned the last time are for the whole CERN and above the current limit of 2250 people. It has now to be dispatched between the different departments.
- Safety Habilitation lctrique => Training course now available on the web courses (in the Training Catalogue).
- Special meeting organized by the DG for the long term scientific plan => Discussions about:
- 2-year break for the 1st SD (2 years for physics but it is 20 months from beam to beam).
=> Triggered 2 proposals:
- ATLAS => 3-year running period between each SD.
- CMS => 1st SD should start in March-April 2013
=> the 3rd SD for HL-LHC is in 2022 instead of 2020 initially foreseen (but we still start the new approved project now).
- HL-LHC kick-off Internal Meeting at the Holiday-Inn Thoiry from 8:30 to 17:00 on FR 15/04/2011.
- The limit for the BE department is 5 fellows / committee, and as there are 2 committees per year (May and November) it leads to a maximum of 10 gets / year. This limit comes from the DG and budget sharing.
- News on the LHC:
- EM gave a talk at the last LBOC on some ecloud measurements during the 2011 LHC scrubbing run.
=> We need to continue the data analysis (on all the equipments). It seems that in all the measurements, a scrubbing effect was observed, which was the goal... The next question is: What was the initial SEY (and R etc.) and which were the final values?
4) The SUSSIX code (EHM): pptx
- SUSSIX: A Computer Code for Frequency Analysis of Non-Linear Betatron Motion. FS mentioned that the SIX in SUSSIX comes from SIXTRACK. It is a FORTRAN program with ~ 150000 lines.
- What is SUSSIX?
- SUSSIX = a program for the post processing of turn-by-turn
betatron oscillation data written by F. Schmidt & R. Bartolini
=> Computes the frequency decomposition of the signal using an interpolated FFT
with time filter to determine the tune to high precision:
- It calculates dynamical quantities: Resonance driving terms, Detuning, Smear
- The data are computed from particle tracking or measured from BPMs.
- The source files, examples and documentation can be found at: http://www.cern.ch/emaclean/.
- How is SUSSIX used?
- sussix.inp file specifies the SUSSIX actions.
- How does SUSSIX work?
- SUSSIX performs a frequency analysis on the provided turn-by-turn data. The time signal with the turn-by-turn data is decomposed into a series of spectral lines:
- The turn by turn data signal is decomposed as a quasi-periodic signal => It works for sinusoidal signals only! But it could be extended to other signals, like exponential ones etc. (see discussion in talk from BS).
- Turn by turn data signal is analysed to give the Amplitude spectrum.
- The line with the largest amplitude is in most cases the fundamental frequency (tune). This can be overridden using the ISTUNE in the sussix.inp file.
- Regular motion means other frequencies are linear combinations of the fundamental tunes.
- At the core of SUSSIX is a Fortran Routine TUNENEWT, developed by the Bologna Group in the mid-nineties (Author: A. Bazzani; tests by R. Bartolini, M. Giovannozzi & E. Todesco):
- TUNENEWT determines the maximum of the amplitude of the FFT.
- SUSSIX uses the TUNENEWT routine to find the fundamental frequencies of the signal. The corresponding harmonic is subtracted and the new signal re-analyzed.
- The spectral lines are ordered, as linear combinations of the tunes, in decreasing amplitude.
- How effective is SUSSIX?
- Use of time window filters improves the effectiveness of the determination of tunes => In SUSSIX a Hanning filter may be applied by selecting the IMETH 1 flag in sussix.inp.
- Applying the filter:
- Broadens the main peak FWHM from 2 / N to 4 / N,
- Reduces the height of side lobes by a factor 1 / N2.
- Scalings:
- FFT scales as 1 / N.
- No time filter Interpolated FFT scales as 1 / N2. Reminder: Interpolated does not mean 0 padding! By interpolating we improve the precision because we know/assume that the signal is sinusoidal. In fact, Massimo mentioned that there is also a closed formula for that and that it could be extended to any known function.
- Hanning filter Interpolated FFT scales as 1 / N4 (without noise) => The 1 / N4 scaling of errors in the interpolated FFT with Hanning filter does not hold when the signal is noisy: it becomes ~ 1 / N2, which is still better than the FFT without noise.
- However the reconstruction of the amplitude and phase of the Fourier coefficient is slightly less precise than for frequency.
- Further limits:
- Due to the finite time signal, the spectral lines have finite FWHM 4 / N (with Hanning filter):
- To resolve the spectral lines they must be separated by at least the FWHM.
- If a spectral line lies within a FWHM, SUSSIX will loop around the correct frequency without properly subtracting the two contributions.
- Observe a cluster of lines about the correct frequency.
- Increasing N will allow closer spectral lines to be distinguished
- The signal is decomposed as a quasi-periodic time series:
- Implies a regular stable orbit.
- In highly chaotic regime or close to resonances cannot produce meaningful results.
- In applications to BPM data SUSSIX is limited by the decoherence of the turn-by-turn oscillations.
- BUT, SUSSIX IS STILL BETTER THAN A FFT!
- Frequency Map analysis is a powerful method of extracting information on the Beam Dynamics => The method relies on being able to very accurately determine the tune. TUNENEWT provides the level of precision necessary for these studies.
- Speeding up SUSSIX:
- SUSSIX represents a bottleneck in the control room: For a single off-momentum measurement, the beta-beat group would be required to analyze 15 sets of ~ 500 BPM signals => > 15minutes spent running SUSSIX per measurement.
- 2 approaches to speeding up the running of SUSSIX have been attempted:
- Parallelization of the FFT using recently released NAG parallel FFT routines => Overheads involved in setting up the parallelization cancelled out the benefits of the parallel FFT.
- Parallelization of the analysis of multiple BPMs:
- At present SUSSIX performs the analysis of each set of BPM data in series.
- Parallelizing the BPM analysis has produced a substantial improvement in the run time of SUSSIX when analysing a large number of BPMs
- Harry Renshall has been working on the parallelization of SUSSIX => Next talk.
- Comment from BS: OP, and in particular Verena Kain, would be interested to use SUSSIX in the CCC but without a Python layer etc... Ideally written in Java...
5) High level parallelization of SUSSIX (Harry Renshall): pptx
- He started with a simple case from Ewen on lxplus.
- The goal was to gain a factor 10 in the speed (for the use of SUSSIX in the CCC for the beta-beat group).
- Use the gnu 77 Fortran compiler g77 with opt = 3 (opt for optimization => Always use opt = 3) => Fortran compilers are very efficient these days.
- Obvious higher level parallelisation is over the individual bpm files which are all independent.
- How the parallelization was done?
- Only few modifications done by Harry => See what is a local variable and declare it to be private for each bpm list.
- The only executable code changes are in the main program.
- The rest are non-executable apart from ordres where all 'write (30,' were changed to 'write (3000+n,'. The core code changes in sussix.s are given in page 3.
- Results
- The factor 10 was reached with ~ 20 threads.
6) Taking advantage of Sussix's sensitivity to reveal higher order mode coupling (BS): pptx
- BS showed 3 examples of application of SUSSIX:
1) Analysis of simulation data from HEADTAIL
- Bunched-beam coherent modes of oscillation => 1st time SUSSIX was used for collective effects (motion of the bunch centre of mass due to the interaction with a broad-band impedance)!
- This is explained in detail in BS' PHD thesis: http://library.epfl.ch/theses/?nr=4585.
- Very good resolution of sidebands by SUSSIX, even for low intensities.
- A very good agreement was reached with MOSES.
2) Analyses of 2 data sets from measurements in the LHC:
- BBQ signals when moving collimators in 2010 => SUSSIX resolution is a bit better than FFT but not so much gain as in simulations.
- Transverse damper data during scrubbing run 2011 => 73 turns of bunch by bunch data available from the application of Riccardo de Maria. Very low frequency resolution of the FFT. Can SUSSIX be trusted with such a small number of turns?
- BS' observations as a SUSSIX' user:
- For the same number of turns, SUSSIX has higher resolution and higher sensitivity than the classical FFT, in particular if the complex signal can be used as input => Very nice tool!
- However, it should not be used in all cases as:
- SUSSIX is slower than the FFT,
- SUSSIX is very slow for large number of turns,
- SUSSIX can give strange results when the motion is exponentially growing (i.e. when the beam is unstable for instance)
- SUSSIX is used heavily in the LHC python-based beta-beat applications for SPS and LHC.
- FFT is present by default or as a library in most codes and languages, while SUSSIX is not (yet).
- Other comments during the meeting:
- Only real signals were used until now.
- We could also use complex signals by using 2 BPMs in the SPS for instance spaced by 90 deg (or in fact by any known phase advance) as the x at the second BPM is linked to the x' of the first => Should/will be used by Nicolo for his transverse impedance localization studies.
- SUSSIX, as programmed now, assumes a sinusoidal signal but it could be extended to any known function, e.g. exponential (see Ewen's talk) => We could do this to try and analyse the HEADTAIL simulation data in the unstable region and compare to MOSES (as the comparison done by BS was only for the stable part). We "just" have to divide by the exponential...
7) Phase advance accuracy (NB): pptx
- NB compared the phase accuracy obtained by 3 different FFT codes (in case of ideal sinusoidal signals with phase advance typical for SPS FODO lattice, and adding Gaussian noise):
1) Ap-FFT (= All-Phase FFT) => NEW METHOD OF ESTIMATION OF PHASE, AMPLITUDE, AND FREQUENCY BASED ON ALL PHASE FFT SPECTRUM ANALYSIS by Huang Xiaohong, Wang Zhaohua and Hou Guoqiang,
2) Matlab FFT,
3) SUSSIX-FFT.
- Without noise => Ap-FFT is the best.
- With 10% Gaussian noise => Ap-FFt still the best for some phases, but of course much less precise than without noise..
- NB than also compared the frequency accuracy => Ap-FFT and Matlba are ~ the same and USSIX is much better than the two in both cases (without and with noise).
- Using Ap-FFT for simulation data increase the accuracy in phase determination for the tracked signals. This permits to gain a better resolution in reconstruction and avoids the peaks forest at the end, which are obtained with Matlab.
8) Actions to be taken for the next meeting
- Old actions.
9) Miscellaneous
- The next (25th) meeting will take place on 20/04/2011 => Agenda:
1) News from the US of the 6D beam-beam lens (FS) + Beam-beam note (FS),
2) Review of IBS: analytic and simulation studies (Alessandro Vivoli).
- See preliminary agendas for the next meetings.
Minutes by E. Metral, 29/04/2011.