]>
Commit | Line | Data |
---|---|---|
b1f776f8 | 1 | |
2 | ||
6ebd8b1a | 3 | \section{The EMCal HLT online chain - Federico} |
b1f776f8 | 4 | |
5 | The EMCal L0 or L1 hardware trigger decisions provide the input | |
6 | for a dedicated on-line event processing chain running on the HLT cluster, | |
7 | where further refinement based on criteria using the full event reconstruction | |
8 | information is performed. | |
9 | In fact, the detector optical link transports the raw data to the | |
10 | Read-Out Receiver Card (RORC) in the local data collector of the data acquisition system, | |
11 | which sends a complete copy of the readout | |
12 | to a set of specialized nodes in the HLT cluster (FEP or Front End Processors). | |
13 | Each FEP node is equipped with RORC cards in analogy to | |
14 | the collector nodes used by the data acquisition. | |
15 | The FEP nodes are physically linked to the detector hardware and reflect the geometrical | |
16 | partitioning of each ALICE sub-system. | |
17 | The 10 full-size super-modules are read out using 2 Read-Out Control Units (RCUs) for a total | |
18 | of 20 optical links running into the HLT FEPs. | |
19 | The reduced-size super-modules were installed prior to the 2012 LHC run and are not discussed in the present report. | |
20 | In addition to the 20 links from the super-module readout, the HLT receives also a copy of the L0/L1 trigger data stream | |
21 | via an additional optical link from the EMCal jet trigger unit (STU) data collector. | |
22 | The different stages of data processing are then performed by the | |
23 | software analysis chain executed on the HLT cluster: | |
24 | a set of general purpose nodes (Computing Nodes or CNs) | |
25 | perform the higher level operations on the data streams which have been | |
26 | already pre-processed on the FEPs at the lower level. | |
27 | The EMCal software components form a specialized sub-chain | |
28 | executed at run time together with all other ALICE sub-systems | |
29 | participating in the HLT event reconstruction. | |
30 | ||
31 | \begin{figure}[ht] | |
32 | \begin{center} | |
3fd21830 | 33 | \includegraphics[width=23pc]{figures/chain-new.pdf} |
b1f776f8 | 34 | \caption{\label{f1} Functional diagram of the EMCal online reconstruction components (signal processing, data structure makers, and clusterizers) shown in green. |
35 | The EMCal chain is fed by the detector raw data. Trigger components are shown in red. EMCal-specific triggers operate on the calorimeter clusters | |
36 | and perform TPC track-matching when needed (electron and jet triggers). Monitoring components are shown in blue and live in a separate monitoring chain. | |
37 | The EMCal triggers are evaluated within the Global Trigger which is aware of the full HLT trigger logic of the other ALICE detectors. | |
38 | } | |
39 | \end{center} | |
40 | \end{figure} | |
41 | ||
42 | The functional units of the EMCal HLT online chain are presented in Figure \ref{f1} where | |
43 | the online reconstruction, monitoring, and trigger components | |
44 | %with their relevant classes | |
45 | are shown together with their relevant data paths. | |
46 | The lower-level EMCal online component ({\it RawAnalyzer}) is fed by the detector front end electronics | |
47 | and performs signal amplitude and timing information extraction. | |
48 | Intermediate components ({\it DigitMaker}) use this information to | |
49 | build the digitized data structures needed for the clusterizer components to operate on the cell signals. | |
50 | Alternatively, the digitized signals can be generated via monte carlo simulations ({\it DigitHandler}). | |
51 | ||
52 | At the top of the EMCal reconstruction chain, the digits are summed by the {\it Clusterizer} component to produce the cluster data structures. | |
53 | The calorimeter clusters are then used to generate the different kinds of EMCal HLT trigger information: | |
54 | a single shower trigger ($\gamma$) with no track matching, an electron trigger using the matching with a corresponding TPC track, | |
55 | and a jet trigger also using the TPC tracks information and the V0 multiplicity dependent threshold. | |
56 | ||
57 | The trigger logic generated by the EMCal chain is evaluated | |
58 | (together with the outputs of the HLT trigger components coming from other ALICE detectors) | |
59 | within the HLT Global Trigger which produces the final high level decision based on the reconstructed event. | |
60 | The ALICE data acquisition system will then discard, accept or tag the event according to the HLT decision. | |
61 | ||
62 | For performance and stability reasons, the full on-line HLT chain contains only analysis | |
63 | and trigger components. On the other hand, monitoring components typically make | |
64 | heavy use of histogramming packages and ESD objects, hence they are kept in a separate chain. | |
65 | The isolation of the monitoring from the reconstruction chain gives additional robustness since | |
66 | a crash in a monitoring component will not affect the reconstruction chain and the data taking. | |
67 | ||
68 | \subsection{Reconstruction components} | |
69 | ||
70 | As shown in Figure \ref{f1} the EMCal HLT analysis chain provides all the | |
71 | necessary components to allow the formation of a trigger | |
72 | decision based on full event reconstruction. The following | |
73 | subsections are devoted to a detailed discussion of each processing | |
74 | stage, starting from the most basic, i.e. signal extraction, to | |
75 | the highest stage: the HLT trigger decision. | |
76 | ||
77 | \subsubsection{RawAnalyzer} | |
78 | The {\it RawAnalyzer} component extracts energy and timing information for each calorimeter cell. | |
79 | Extraction methods implemented in the offline code (AliRoot) typically | |
80 | use least squares fitting algorithms, and cannot be used in online processing for | |
81 | performance reasons. | |
82 | Conversely, the HLT signal extraction is done without need of fitting | |
83 | using two possible extraction methods. The first method, referred to as {\it kCrude}, | |
84 | simply produces an amplitude using the difference between the maximum and the minimum values | |
85 | of the digitized time samples and associates the time bin of the maximum as the signal arrival time. | |
86 | The {\it kCrude} method was used during the 2011 data taking: it has the advantage | |
87 | of being extremely fast and fully robust since no complex algorithms are used. On the other hand, | |
88 | it produces a less accurate result than the processing of the full signal shape. | |
89 | An alternative method ({\it kPeakFinder}) evaluates the amplitude | |
90 | and peak position as a weighted sum of the digitized samples. This approach is | |
91 | not as fast as {\it kCrude} but is a few hundred times faster than least squares fitting. | |
92 | ||
93 | \subsubsection{DigitMaker} | |
94 | The {\it DigitMaker} component essentially transforms the raw cell signal amplitudes | |
95 | produced by the {\it RawAnalyzer} into digit structures by processing the cell coordinates | |
96 | and by the application of dead channel maps and the appropriate gain factors (low and high-gain). | |
97 | ||
98 | \subsubsection{Clusterizer} | |
99 | The {\it Clusterizer} component merges individual signals (digits) of adjacent cells | |
100 | into structures called clusters. | |
101 | At transverse momenta $p_T>1$~GeV/c most of the clusters are associated to electromagnetic | |
102 | showers in EMCal from $\pi^0$ and $\eta$ mesons decays. | |
103 | Other sources of electromagnetic showers are direct photons | |
104 | and electrons from semi-leptonic decays of $c$ and $b$ hadrons. | |
105 | Since the typical cluster size in the EMCal can vary according to the detector occupancy due to shower | |
106 | overlap effects, which are much different for {\it pp} and heavy-ion collisions, | |
107 | clustering algorithms with and without a cutoff on the shower size are available (both in offline and in the HLT) | |
108 | to optimize the cluster reconstruction for the different cases. | |
109 | Events originating from {\it pp} collisions tends to generate | |
110 | smaller, spherical and well-separated clusters in the EMCal, at least up to 10 GeV/c. | |
111 | At higher transverse momenta, overlapping of the showers requires a shape analysis | |
112 | to extract the single shower energy. | |
113 | Above 30 GeV/c the reconstruction can be performed only with more sophisticated | |
114 | algorithms such as isolation cuts to identify direct photons. | |
115 | ||
116 | ||
117 | \begin{figure}[ht] | |
118 | \begin{minipage}{16pc} | |
119 | %\includegraphics[width=18pc]{clus-comp.eps} | |
3fd21830 | 120 | \includegraphics[width=16pc]{figures/emcal_cluster_match_nxn.pdf} |
b1f776f8 | 121 | \caption{\label{f5} Reconstruction efficiency for the $N\times N$ algorithm (cutoff) in offline and HLT. The notation $(A) \rightarrow (B)$ indicates |
122 | the fraction of clusters found using method A that are also found using method B (data from run 154787, period LHC11c).} | |
123 | \end{minipage}\hspace{2pc} | |
124 | \begin{minipage}{16pc} | |
3fd21830 | 125 | \includegraphics[width=16pc]{figures/emcal_cluster_match_v1.pdf} |
b1f776f8 | 126 | \caption{\label{f6}Reconstruction efficiency for the $V1$ algorithms (no cutoff) in offline and HLT. The notation $(A) \rightarrow (B)$ indicates |
127 | the fraction of clusters found using method A that are also found using method B (data from run 154787, period LHC11c).} | |
128 | \end{minipage} | |
129 | \end{figure} | |
130 | ||
131 | The identification of an isolated single electromagnetic cluster in the EMCal can be performed using different | |
132 | strategies: summing up all the neighboring cells around a seed-cell over threshold until no more cells are found | |
133 | or adding up cells around the seed until the number of clustered cells reaches the predefined cutoff value. | |
134 | ||
135 | The first approach is more suitable for an accurate reconstruction. A further improvement to this | |
136 | clustering algorithm would be the ability to unfold overlapping clusters as generated from the | |
137 | photonic decay of high-energy neutral mesons, however this procedure usually requires computing intensive fitting algorithms. | |
138 | ||
139 | Such performance penalty must be avoided in the online reconstruction so the cutoff | |
140 | technique is preferred. In the EMCal HLT reconstruction a cutoff of 9 cells is used (according to the | |
141 | geometrical granularity of the single cell size), so the clusterization is performed into a | |
142 | square of $3\times3$ cells. The cutoff and non-cutoff algorithms are referred to as $N\times N$ and $V1$, respectively. | |
143 | ||
144 | In {\it pp} collisions the response of the two methods is very similar since the majority of clusters are well separated, while | |
145 | in {\it PbPb} collisions, especially in central events, the high particle multiplicity requires the use of the cutoff (or unfolding in offline) | |
146 | to disentangle the cluster signals from the the underlying event to avoid the generation of artificially large clusters. | |
147 | ||
148 | The quality of the EMCal online clusterizer algorithms implemented in the HLT chain were checked against offline, | |
149 | as shown in Figures \ref{f5} and \ref{f6} where it can be seen that the performance is in a reasonable agreement in all cases. | |
150 | The low point at 1.25 GeV is due to bad towers, which are assigned an energy of 1 GeV. | |
151 | %The HLT clusterizer does not have the capability to remove bad clusters. | |
152 | Bad clusters are removed in later stages of the analysis, but that is not yet reflected in Figures \ref{f5} and \ref{f6}. | |
153 | This effect leads to an excess of clusters that are found by the HLT clusterizer, but not by the offline clusterizer.\\ | |
154 | ||
155 | Since the EMCal HLT reconstruction is mainly targeted for triggering, a small penalty in the accuracy of the energy reconstruction of the clusters is accepted | |
156 | as a trade off in favor of faster performance, and for this reason the cutoff clustering method was used, especially in {\it PbPb} collisions. | |
157 | ||
158 | \subsection{Trigger components} | |
159 | ||
160 | The online HLT chain is capable of producing trigger decisions based on full | |
161 | event reconstruction. In terms of EMCal event rejection the following relevant trigger observables | |
162 | have been implemented: | |
163 | ||
164 | \begin{itemize} | |
165 | \item neutral cluster trigger | |
166 | \item electron and jet trigger | |
167 | \end{itemize} | |
168 | ||
169 | ||
170 | \subsubsection{Cluster trigger} | |
171 | The single shower triggering mode is primarily targeted to trigger on photons and neutral mesons. | |
172 | In all collision systems, the high level trigger post-filtering can improve | |
173 | the hardware L0 and L1 trigger response by using the current bad channels map information | |
174 | and calibration factors (which could be recomputed directly in the HLT). | |
175 | ||
176 | \subsubsection{Electron trigger} | |
177 | For this trigger the cluster information reconstructed online by the EMCal HLT analysis | |
178 | chain is combined with the central barrel tracking information to produce complex event selection | |
179 | as a single electron trigger (matching of one extrapolated track with an EMCal cluster. | |
180 | % TCA - this shouldn't be in this section, if it is included... | |
181 | %, and requiring a match between the cluster energy and the track momentum) and a full jet trigger | |
182 | %(matching of multiple tracks with a jet patch in the EMCal). | |
183 | Performance and accuracy studies of the track matching component developed for this purpose | |
184 | have been done using simulated and real data taken during the 2011 LHC running period. | |
185 | Results are shown in Figures \ref{f7} and \ref{f8} where the cluster - track residuals | |
186 | in azimuth and pseudo-rapidity units are to be compared with a calorimeter cell size of | |
187 | $0.014\times 0.014$. | |
188 | ||
189 | \begin{figure}[hb] | |
190 | \begin{minipage}{16pc} | |
3fd21830 | 191 | \includegraphics[width=15.5pc]{figures/Fig4dPhi_performance.pdf} |
b1f776f8 | 192 | \caption{\label{f7} |
193 | Distribution of the residuals in azimuth ($\Delta\phi$) for the EMCal cluster and central barrel tracks | |
194 | obtained using the HLT online chain for run 154787 (LHC11c), ~ 70 k events reconstructed. | |
195 | } | |
196 | \end{minipage}\hspace{2pc} | |
197 | \begin{minipage}{16pc} | |
3fd21830 | 198 | \includegraphics[width=16pc]{figures/Fig5dEta_performance.pdf} |
b1f776f8 | 199 | \caption{\label{f8} |
200 | Distribution of the residuals in pseudo-rapidity ($\Delta\eta$) for the EMCal cluster and central barrel tracks | |
201 | obtained using the HLT online chain for run 154787 (LHC11c), ~ 70 k events reconstructed. | |
202 | } | |
203 | \end{minipage} | |
204 | \end{figure} | |
205 | ||
206 | ||
207 | In addition to the extrapolation of the track from the central barrel | |
208 | to the EMCal interaction plane and the matching with a compatible nearby cluster, | |
209 | the electron trigger component must finally perform particle identification | |
210 | to issue a trigger decision. The selection of electron candidates is done | |
211 | using the $E/pc$ information where the energy is measured from the | |
212 | EMCal cluster and the momentum from the central barrel track. | |
213 | The trigger component is initialized with default values | |
214 | for the cut of $0.8< E/pc <1.3$. The default cuts are stored in the HLT | |
215 | conditions database and can be overridden via command line arguments | |
216 | at configuration time (usually at start of run). | |
217 | ||
218 | The performance of the electron trigger was studied using {\it pp} minimum | |
219 | bias data at 7 TeV with embedded $J/\Psi$ events. | |
220 | Figure \ref{f9} shows the good agreement of the $E/pc$ distributions | |
221 | obtained with the track extrapolation - cluster matching | |
222 | performed using the online algorithms compared to the ESD-based tracking (red). | |
223 | ||
224 | \begin{figure}[ht] | |
3fd21830 | 225 | \includegraphics[width=24pc]{figures/Fig6HLTEoverP_performance.pdf} |
b1f776f8 | 226 | \begin{center} |
227 | \caption{\label{f9} | |
228 | $E/pc$ distributions obtained with the track extrapolation - cluster matching | |
229 | via the online algorithms compared to the ESD-based tracking (red).} | |
230 | \end{center} | |
231 | \end{figure} | |
232 | ||
233 | To determine the possible improvement of the event selection | |
234 | for electrons with energies above 1~GeV, AliRoot simulations of the HLT chain using LHC11b10a {\it pp} minimum bias data | |
235 | at 2.76 GeV and the EMCal full geometry (10 super-modules) have been used. These studies | |
236 | have shown that at least a factor 5 to 10 in event selection can be gained compared to the single shower trigger, as shown in Figure \ref{f10}. | |
237 | ||
238 | \begin{figure}[ht] | |
3fd21830 | 239 | \includegraphics[width=24pc]{figures/Fig7Events_performance.pdf} |
b1f776f8 | 240 | \begin{center} |
241 | \caption{\label{f10} | |
242 | Improvement in the event selection for $E_{e^-}>$~1~GeV from AliRoot simulation (anchor to LHC11b10a) with minimum bias {\it pp} at $\sqrt{s}=2.76$~TeV (EMCal full geometry). | |
243 | The red points are obtained with the requirement of one hit in one of the silicon pixel (SPD) layers to reject a higher fraction of photon conversions. | |
244 | } | |
245 | \end{center} | |
246 | \end{figure} | |
247 | ||
248 | ||
249 | \subsubsection{Jet trigger} | |
250 | ||
251 | The EMCal online jet trigger component was developed to provide | |
252 | an unbiased jet sample by refining the hardware L1 trigger decisions. | |
253 | In fact, the HLT post-processing can produce a sharper turn on curve | |
254 | using the track matching capabilities of the online reconstruction chain. | |
255 | In addition, a more accurate definition of the jet area than the one provided by the hardware L1 jet patch, | |
256 | can be obtained choosing a jet cone based on the jet direction calculated online. | |
257 | The combination of the hadronic and electromagnetic energy provides a measurement of | |
258 | the total energy of the jet by matching the tracks identified as part of the jet with | |
259 | the corresponding EMCal neutral energy. | |
260 | ||
261 | The use of the HLT jet trigger also allows a better characterization | |
262 | of the trigger response as a function of the centrality dependent threshold | |
263 | by re-processing the information from the V0 detector directly in HLT. | |
264 | ||
265 | Performance considerations, due to the high particle multiplicity | |
266 | in {\it PbPb} collisions, impose that the track extrapolation is done only geometrically | |
267 | without taking into account multiple scattering effects | |
268 | introduced by the material budget in front of the EMCal. | |
269 | The pure geometrical extrapolation accounts for a speedup factor of 20 in the | |
270 | execution of the track matcher component with respect to the | |
271 | full-fledged track extrapolation used in {\it pp} collisions. | |
272 | ||
273 | The identification of the jet tracks is performed using the anti-$k_T$ | |
274 | jet finder provided by the FastJet package. | |
275 | ||
276 | The EMCal jet trigger was only partially tested during the 2011 data taking period | |
277 | and will be fully commissioned for the LHC {\it pPb} run period in 2012. | |
278 | ||
279 | \subsection{Monitoring components} | |
280 | ||
281 | The role of the EMCal HLT reconstruction in {\it pp} collisions is targeted mainly on | |
282 | the monitoring functions since the expected event sizes are small enough for | |
283 | the complete collision event to be fully transferred to permanent storage. | |
284 | \begin{figure}[h] | |
3fd21830 | 285 | \includegraphics[width=26pc]{figures/all-monitor.pdf} |
b1f776f8 | 286 | \begin{center} |
287 | \caption{\label{f2} Output from the EMCal HLT monitoring component. Top left: cluster energy spectra as a function of the reconstructed cluster energy; | |
288 | bottom left: cluster position in $\eta$ and $\phi$ coordinates; bottom right: cluster time distribution; top right: number of cells per cluster vs cluster | |
289 | energy. LHC11b period, $\sqrt{s}=7$ TeV {\it pp} data, 10~kEvent analyzed.} | |
290 | \end{center} | |
291 | \end{figure} | |
292 | ||
293 | In this respect, two monitoring components have been developed and deployed in the online chain. | |
294 | The first component currently monitors reconstructed quantities, such as the cluster energy spectra and timing, | |
295 | the cluster position in the $\eta$ and $\phi$ coordinates, and the number of cells per cluster | |
296 | as a function of the cluster reconstructed energy as shown in Figure \ref{f2}. | |
297 | ||
298 | \begin{figure}[h] | |
3fd21830 | 299 | \includegraphics[width=24pc]{figures/l0clusters.pdf} |
b1f776f8 | 300 | \begin{center} |
301 | \caption{\label{f3} Energy spectrum for all clusters reconstructed by the EMCal (black points) superposed with the triggered cluster spectrum | |
302 | (i.e. clusters reconstructed which also carry the L0 hardware trigger bit set, red points). } | |
303 | \end{center} | |
304 | \end{figure} | |
305 | ||
306 | The second component re-evaluates the EMCal hardware trigger decisions by recalculating | |
307 | the cluster energy spectrum for all the clusters with the L0 trigger bit set as shown | |
308 | in Figure \ref{f3}. The L0 turn on curve can then be calculated online as the ratio between | |
309 | the triggered and the reconstructed cluster spectra and monitored for the specific run. | |
310 | ||
311 | ||
312 | No recalculation of hardware L1 trigger primitives was possible | |
313 | during the 2011 data taking since the optical link from the EMCal L1 trigger unit | |
314 | could only installed during the 2011-2012 winter shutdown of the LHC hence | |
315 | the software development for the L1 trigger monitoring is still underway. | |
316 |