L1 Triggering on High-Granularity Information at the HL-LHC
Fedor Ratnikov
Alexander Bazilevsky
Ariel G. Schwartzman
Round 1
Reviewer 1 Report
Dear author,
I found your document an interesting reading about the architecture and algorithms for the CMS HGCAL trigger.
This type of document is intrinsically technical, however in some points I found the discussion a bit too difficult to follow for someone that is not working on the HGCAL project, since a few terminologies are not defined and need to be deduced from the context. The document would benefit of very few additions/clarifications to guide the reader.
I am listing below a few questions and suggestions. Being the proceeding of a conference talk, these points are not meant to be a review of the HGCAL developments themselves, but as possible unclear points for a reader that could be clarified in the document.
- l. 4: suggest to quote the actual 12.5 us latency here
- l 5 vs l. 33 vs l. 35: sometimes "Level 1", "Level-1", and "L1" are used, I suggest to use a uniform expression
- l. 36 (also l. 101-103): trigger towers and trigger clusters have a very important role in the document but they are never defined, so it is not clear to the reader how they differ and what is the physical property that the HGCAL trigger captures through them (e.g. single particles vs spread/fixed size energy deposits?). I suggest to add one sentence here to clarify these aspects.
- l. 41: Figure 2 should likely be reference instead
- l. 42: high/low density modules are mentioned but never introduced - one can guess from the figure that the referred density is per number of sensor pads, but it would be good to clarify
- l. 69: is this clustering done separately for each layer, or with merged layers? I would assume the first since 14 boards serve a single sector, but the clustering in r/z suggests an angular grouping criterion. I would suggest clarifying at which level these operations are performed, also to better understand the interplay with respect to the stage-2 clustering
- paragraph 2.1: time resolution is mentioned (l. 43) but it's not clear whether this information is available to the L1 trigger at all
- Fig. 5: for some cases (esp. for eta > 2.2) the efficiency without truncation is lower than the one with truncation, which seems counterintuitive. Could the reason be clarified in the text?
- Fig 5 and Fig. 7: does the HGCAL trigger primitive generation stop somewhere between eta = 2.8 and 2.9 even if the coverage is up to eta = 3 (from Fig. 1)? Maybe the ranges could be mentioned in the text
- ll 115-125: all the fixed thresholds are significantly tighter than the area-normalized one, so why not just area-normalizing higher thresholds? Also, shouldn't lower thresholds result in any case in higher efficiency at the cost of higher multiplicity, unless some truncation is imposed?
- l 143: possibly capitalize Identification for the acronym
Author Response
Dear referee,
Many thanks for the useful comments and suggestions. Please find below some answers.
Best regards,
Louis Portalès
- l. 4: suggest to quote the actual 12.5 us latency here
Done.
- l 5 vs l. 33 vs l. 35: sometimes "Level 1", "Level-1", and "L1" are used, I suggest to use a uniform expression
“Level-1” is now used in abstract and L.33, with correspondence to “L1” mentioned. L1 is used everywhere else.
- l. 36 (also l. 101-103): trigger towers and trigger clusters have a very important role in the document but they are never defined, so it is not clear to the reader how they differ and what is the physical property that the HGCAL trigger captures through them (e.g. single particles vs spread/fixed size energy deposits?). I suggest to add one sentence here to clarify these aspects.
The trigger clusters provide somewhat granular information on shower shapes that can be used (e.g.) for the identification of electrons/photons vs neutral hadrons. Towers are coarser objects that are primarily used in the definition of hadronic jets, in a format consistent with the barrel calorimeter trigger primitives. A brief explanation has been added to the text
- l. 41: Figure 2 should likely be reference instead
Indeed. This is now fixed.
- l. 42: high/low density modules are mentioned but never introduced - one can guess from the figure that the referred density is per number of sensor pads, but it would be good to clarify
This does indeed refer to the number of sensors in the modules. Some clarifications were added in the introduction.
- l. 69: is this clustering done separately for each layer, or with merged layers? I would assume the first since 14 boards serve a single sector, but the clustering in r/z suggests an angular grouping criterion. I would suggest clarifying at which level these operations are performed, also to better understand the interplay with respect to the stage-2 clustering.
No clustering is applied during stage-1, so I suppose you are referring to the grouping of TCs in (r/z,phi) bins? If so, the grouping indeed involves an angular criterion with r/z, and each bin corresponds to a specific 3D region of the detector. The corresponding bullet has been rephrased slightly to make it clearer.
- paragraph 2.1: time resolution is mentioned (l. 43) but it's not clear whether this information is available to the L1 trigger at all
Due to the very limited bandwidth, only information on the TC energies is propagated through the TPG blocks. The timing information is only propagated to the DAQ. The mention of time of arrival has been removed to avoid confusion.
- Fig. 5: for some cases (esp. for eta > 2.2) the efficiency without truncation is lower than the one with truncation, which seems counterintuitive. Could the reason be clarified in the text?
The differences between the four curves is at the per-mil level, and are likely due to statistical fluctuations. The text has been changed from “has little impact” to “has no impact” to better fit this point.
- Fig 5 and Fig. 7: does the HGCAL trigger primitive generation stop somewhere between eta = 2.8 and 2.9 even if the coverage is up to eta = 3 (from Fig. 1)? Maybe the ranges could be mentioned in the text
The TP generation will indeed cover the full acceptance, up to |eta|=3. The range used in the plot was chosen for this study only to focus on clusters fully reconstructed (offline) within HGCAL, avoiding border effects. This information is now added to the legends.
- ll 115-125: all the fixed thresholds are significantly tighter than the area-normalized one, so why not just area-normalizing higher thresholds?
The area-normalized threshold is only looser than the fixed thresholds that are studied for larger |eta|. At low |eta| the area-normalization included on top of the baseline 8.5mipT value actually makes it much tighter, hence yielding the lower efficiency seen on Fig.7a. This significant |eta| dependence is the effect that we want to get rid of by introducing fixed thresholds.
- Also, shouldn't lower thresholds result in any case in higher efficiency at the cost of higher multiplicity, unless some truncation is imposed?
Indeed, this is what is seen in Fig. 7a. I am not sure if this is part of what triggered the question, but there was a typo in the text, where it was mentioned that thresholds “larger than 20 mipT” are preferred, when it should have said “lower than”. This is now fixed.
- l 143: possibly capitalize Identification for the acronym
Done
Reviewer 2 Report
This is a solid, well-written paper.
Below are several suggestions that could improve the quality and value of the paper.
- Figures are in raster format. Taking into account the presence of tiny labels, the available resolution is often not enough to clearly read them. Suggest to re-make figures in vector format
- l41: reference to Fig. 2, not 3
- Chapter 2: there are detailed technical descriptions of different trigger components, however, the basic physics concept is not highlighted. Which components process individual layers, which components combine layers into 3D sectors, which kind of physics proto-objects are processed/produced by Stage 1 and Stage 2
- Figure 4 needs more explanation for a more generic reader
- l107-108: Why does smearing histogram improves peak finding capability? Some reference or explanation is required
- l 110-111: Is it correct that the cluster has ~1radian in radius? This is huge! was this number optimized in any way?
- Fig. 7a is a bit cryptic. The relation between black and colour lines needs an explanation
- l.152: citing 300 GeV as a difference in threshold doesn't look like apples to apples comparison. They are different algorithms with different relations between algorithm threshold and efficiency curves. A more direct number would be an effective energy threshold for different algorithms at some reasonable working point for efficiency. It means a horizontal line corresponding to the reference efficiency, and offsets for different algorithms along this line
- 3.1-3.3: these sections miss technical details of how pretty complicated algorithms under discussion fit into Correlator Trigger and satisfy L1 trigger latency requirements. Are algorithms simplified and serialized etc?
Author Response
Dear referee,
Thanks a lot for the useful comments and suggestions. You will find some answers to them below.
Best regards,
Louis Portalès
- Figures are in raster format. Taking into account the presence of tiny labels, the available resolution is often not enough to clearly read them. Suggest to re-make figures in vector format
Changed figures to pdf
- l41: reference to Fig. 2, not 3
Done
- Chapter 2: there are detailed technical descriptions of different trigger components, however, the basic physics concept is not highlighted. Which components process individual layers, which components combine layers into 3D sectors, which kind of physics proto-objects are processed/produced by Stage 1 and Stage 2
None of the algorithm blocks is looking at individual layers, but rather at regions defined in the (r/z,phi) plane. As mentioned at the start of the paragraph 2.1, the Stage-1 FPGA receive TC information, and outputs time-multiplexed collections of trigger cells, and “partial” trigger towers. These are the input of the Stage-2 blocks. Through the “clustering path”; the TC collection is used to form 3D trigger clusters, and through the “Tower energies path”, the partial trigger towers are summed to form the trigger towers. These 3D clusters and trigger towers are the object that are sent to the L1 trigger.
- Figure 4 needs more explanation for a more generic reader
The figure and legend have been slightly modified for more clarity.
- l107-108: Why does smearing histogram improves peak finding capability? Some reference or explanation is required
The smearing reduces the impact of fluctuations, and the probability to reconstruct a cluster centred on it. The sentence has been rephrased.
- l 110-111: Is it correct that the cluster has ~1radian in radius? This is huge! was this number optimized in any way?
This statement was not factual and is now corrected. The clustering radius is indeed much smaller, and defined in a layer-dependant way.
- Fig. 7a is a bit cryptic. The relation between black and colour lines needs an explanation
Some details were added to the text, and the legend now includes a short description of the different curves in the plots.
- l.152: citing 300 GeV as a difference in threshold doesn't look like apples to apples comparison. They are different algorithms with different relations between algorithm threshold and efficiency curves. A more direct number would be an effective energy threshold for different algorithms at some reasonable working point for efficiency. It means a horizontal line corresponding to the reference efficiency, and offsets for different algorithms along this line
I do believe however that the comparison shown is meaningful, but was missing the comment that on top of having, for the same L1 rate, a lower L1 threshold on HT, a sharper turn-on is achieved. The figure description in the text now includes this point. This is of course just a qualitative comparison, and the one you are suggesting would indeed be more direct and precise, but I am afraid that it would be difficult, time-wise, to obtain such results for this proceeding.
- 3.1-3.3: these sections miss technical details of how pretty complicated algorithms under discussion fit into Correlator Trigger and satisfy L1 trigger latency requirements. Are algorithms simplified and serialized etc
A mention has been added in the chapter introduction of the “upgraded hardware, allowing much ample room than during Run~2 and 3 to implement complex L1 algorithm, using topological information on the events and advanced Machine-Learning algorithms, adapted to be implemented directly in the L1 trigger firmware”
The technical description of how the more complex algorithms (especially PFLow and PUPPI) are implemented in firmware seemed to go beyond the scope of this proceeding, that is meant to focus mostly on the role of HGCAL in the L1 trigger chain.
Reviewer 3 Report
A few minor suggestions:
Line 21: "5-7.5 x 10^34" => "(5-7.5) x 10^34"
Line 41: "Figure 3" => "Figure 2"
Fig.6 caption: "Stage 1" => "Stage 2"
Line 110: "Delta R^2 =1" : what units?
Line 120: "mipT later on in the text" => "referred to as mipT later on in the text"
Author Response
Dear referee,
Thanks a lot for the corrections and questions. You'll find some answers to them below.
Best regards,
Louis Portalès
Line 21: "5-7.5 x 10^34" => "(5-7.5) x 10^34"
Done
Line 41: "Figure 3" => "Figure 2"
Done
Fig.6 caption: "Stage 1" => "Stage 2"
Done
Line 110: "Delta R^2 =1" : what units?
Delta R is unitless in that context, as it is defined in the (x/z,y/z) plane. The statement was however incorrect, and has been fixed: the clustering radius is layer-dependant, and much smaller than 1.
Line 120: "mipT later on in the text" => "referred to as mipT later on in the text"
Done
Reviewer 4 Report
This article describes studies of jet trigger architecture design and performance of the high granularity calorimeter (HGCAL) upgrade of the CMS experiment. The complexity of the HL-LHC event and the incorporation of HGCAL at the trigger level to handle the challenge of pileup require innovations to solve a large number of technical challenges that the authors describe. This work is extremely interesting and will play a significant role in delivering science at the HL-LHC.
I have a few minor questions/comments:
- Figures 3 and and 6 have a question mark inside one box. Is this intentional or a typo?
- Figure 4: Is there a way to highlight better the 5 outputs?
- Figure 9: are the 2 figures consistent with each other? The plot on the right suggest a very good calorimeter jet resolution (almost as good as PUPPI). However, when multiple jets are combined into an HT trigger, the difference between calo and PUPPI is enormous. Where do these differences come from?
Author Response
Dear referee,
Thanks a lot for the useful questions. You will find answers to them below.
Best regards,
Louis Portalès
- Figures 3 and and 6 have a question mark inside one box. Is this intentional or a typo?
The question marks are intentional, and point to possible change to the algorithm design that could already be implemented, but are in principle not required. The figures legends now mention that the “current” algorithm blocks are shown.
- Figure 4: Is there a way to highlight better the 5 outputs?
The figure and legend has been slightly modified to make it clearer, with arrows now indicating the inputs/outputs.
- Figure 9: are the 2 figures consistent with each other? The plot on the right suggest a very good calorimeter jet resolution (almost as good as PUPPI). However, when multiple jets are combined into an HT trigger, the difference between calo and PUPPI is enormous. Where do these differences come from?
A key difference between the two figures is the |eta| coverage. In Figure 9a, the full |eta|<2.4 is covered, hence including barrel jets, with poorer calorimeter resolution and therefore a much larger gain from PUPPI, that uses also tracking information. Figure 9b focuses on the endcaps only (1.5<|eta|<2.4), and illustrates the nice performance expected from the HGCAL alone.
