LAT PROJECT DOCUMENT CHANGE NOTICE (DCN) SHEET 1 OF 1

Transcription

1 DCN No. LAT-XR LAT PROJECT DOCUMENT CHANGE NOTICE (DCN) SHEET 1 OF 1 ORIGINATOR: J.J. Russell PHONE: DATE: 4/30/03 CHANGE TITLE: DCN for LAT Global Trigger and ACD Hitmaps ORG.: DOCUMENT NUMBER TITLE NEW REV. LAT-TD LAT Global Trigger and ACD Hitmaps 01 CHANGE DESCRIPTION (FROM/TO): Initial release REASON FOR CHANGE: ACTION TAKEN: Change(s) included in new release DCN attached to document(s), changes to be included in next revision Other (specify): DISPOSITION OF HARDWARE (IDENTIFY SERIAL NUMBERS): DCN DISTRIBUTION: No hardware affected (record change only) List S/Ns which comply already: List S/Ns to be reworked or scrapped: List S/Ns to be built with this change: List S/Ns to be retested per this change: SAFETY, COST, SCHEDULE, REQUIREMENTS IMPACT? YES NO If yes, CCB approval is required. Enter change request number: APPROVALS DATE OTHER APPROVALS (specify): DATE ORIGINATOR: J.J. Russell (signature on file) 4/30/03 ORG. MANAGER: Gunther Haller (signature on file) 4/30/03 A.P. Waite (signature on file) 4/30/03 DCC RELEASE: Natalie Cramar (signature on file) 4/30/03 Doc. Control Level: Subsystem LAT IPO GLAST Project DCN No: LAT-XR FORM # LAT-FS

2 Page 1 of 13 Document # Date Effective LAT-TD January 2002 Prepared by(s) Supersedes J.J.Russell None GLAST LAT TECHNICAL DOCUMENT Subsystem/Office Electronics Subsystem Document Title Gamma-ray Large Area Space Telescope (GLAST) Large Area Telescope (LAT)

3 Page 2 of 13 CHANGE HISTORY LOG Revision Effective Date Description of Changes 0 17 Jan 2002 Original

4 Page 3 of 13 Table Of Contents Gamma-ray Large Area Space Telescope (GLAST)...1 Large Area Telescope (LAT) Scope Caveat 4 2 Introduction Alignment of the Trigger Signals Forming the Trigger Window Window Opening Window Closing Closing on the Down Transition of the Signal that Open the Window Closing on the Down Transition of Any Signal that Could Have Opened the Window Window Closes at a Fixed Time after Window Open The Trigger Vector ACD Hit-Map Values and their Usage Arguments For and Against Latching on the FREE Board/Extending ACD Signal Other Consequences of Capturing the VETO Hit-Map on the GLT Forming the Trigger Control Signals...13 Figures Figure 1 Nominal Subsystem Timing... 5 Figure 2 Aligned Subsystem Timing... 6 Figure 3 Timing Diagram Illustrating Erroneous Window Closure... 7 Figure 4 Timing Diagram Illustrating Optimal WIndow Closure... 8 Figure 5 Latching the Trigger Vector... 9

5 Page 4 of 13 1 Scope This document started off as a document to describe the timing requirements to the ACD group. It was quickly realized that many of these requirements come about as a direct consequence of the trigger design. These decisions also have an impact on the physics. The impact on the physics is far too subtle to have been addressed by Level 2 or Level 3 requirements. The trigger section of this document dovetails with LAT-SS-286, The Conceptual Design of the Global Trigger, essentially defining the input signals used by LAT-SS-286. LAT-SS- 284, Trigger Level 4 Subsystem Specification provides the technical requirements and constraints within which the trigger must function. 1.1 Caveat I am under no misconception that I understand all these issues. But I also believe that no one else understands the problem in the whole. Therefore, I thought I would write down my best guess at the possible solutions and advocate the one that I feel is the most optimal. This is meant to be a discussion document, not a design document. I am not religious about the choices I ve made. The only point I would make is that in reviewing the ACD design, many times engineering choices where resolved by the picking the one that gave the highest veto efficiency. While this indeed may be the best choice, I had a gut reaction that the consequences of an over-zealous veto and its impact of gamma efficiency through false vetoes were not adequately considered. 2 Introduction The 4 issues to be examined here are Alignment of the trigger signals Making coincidences of the trigger signals Where to capture the ACD veto hit-map, ACD FREE board or GLT o Sub-issue of false vetoes Formation of the trigger vector The first two items are mandatory. In order to form the LAT trigger, these must be accomplished. Since the ACD is such an integral part of the trigger and visa versa, each influences the other s design. Where to capture the ACD veto hit-map is an issue that has only recently come up. Once all the ACD signals became available on the GLT, an opportunity for overall system simplification became available. In addition to possible system simplification, there is, arguably, a case that the capturing of the ACD veto hit-map at the trigger level may also provide better information.

6 Page 5 of 13 3 Alignment of the Trigger Signals Each subsystem contributes a number of signals to be used in forming the LAT trigger. These signals all have a latency with respect to the exact event time, hereafter referred to as T0, and natural jitter associated with them. Signals must be stretched by their jitter time so that there is always a point after T0 when a signal is guaranteed to be present (strictly speaking the stretch should be little more than the jitter time so that this guaranteed signal spot has finite width). The rest of this document will assume the timing characteristics in the table below (not meant to be accurate, just representative): Latency (nsec) Jitter (nsec) Stretch (nsec) TKR signal (TKRS) ACD signal (ACDS) CAL signal (CALS) Table 1 Subsystem Latency and Jitter Times The following diagram depicts these numbers graphically. The light gray area shows the range of possible rise times for a particular signal and the dark gray area shows the range of possible fall times. To demonstrate that the fall time is deterministic given the rise time, the dark lines show a typical signal for each of the TKRS, ACDS and CALS. TKRS ACDS CALS Figure 1 Nominal Subsystem Timing The immediate problem is that the natural latencies result in non-alignment. Introducing a delay into the ACDS and CALS signals can cure this. Delaying the ACDS signal by 1050 nsec and the CALS signal by 500 nsec results in a sweet spot (at 1200 nsec) where signals are bound to coincide regardless of jitter times:

7 Page 6 of 13 TKRS ACDS CALS Figure 2 Aligned Subsystem Timing The consequence is that every trigger signal or group of signals with similar latencies must have an associated delay register with enough range to accommodate this. Practically speaking the TKR determines time early for the sweet spot and other subsystems must delay up to 1.25usecs to match. To be safe, this number should be at least 1.60usecs. 4 Forming the Trigger Window The LAT trigger signals arrive asynchronously. To form a trigger, one must define a time when the trigger signals will be looked at. This is referred to as the trigger window. The opening and closing of the window must be defined. 4.1 Window Opening The opening of the trigger window is defined by the up transition of any trigger signal that contributes to the trigger decision. This definition may be safely modified to include only signals that may contribute to a positive trigger decision. Intuitively this makes sense, why start a trigger on a signal that cannot possibly lead to a trigger? This issue is important because initiating a trigger sequence will result in deadtime of around nsec. For example, if any and all of the ACD signals where allowed to start a trigger sequence, this would result in 500 nsec * 100 KHz = 5% deadtime (assuming roughly 1KHz rate from each tile.) 4.2 Window Closing The closing of the trigger window can be defined by any of the following conditions The down transition of the signal that opened the window The down transition of any signal that could have opened the window A fixed time delay Each of these will be examined for their advantages and disadvantages Closing on the Down Transition of the Signal that Open the Window The advantage of this method is that the trigger integration period is naturally set by the

8 Page 7 of 13 initiating signal s jitter time. For example, a trigger window opened by the TKR must last at least the 500 nsec. TKR jitter time. However, a trigger window opened by a signal with a smaller jitter need only be opened for that time. Unfortunately the only narrow signals, the ACD, usually do not start a trigger sequence. The ACD CNO signal would fall in the category of being a narrow signal that could start a trigger, but, given the low rate and the CNO usage, it should not be used to decide this issue. The two most likely initiators, the TKR and CAL are both in the nsec range. The disadvantage of this method is arbitrating when more than one signal initiates the window opening. This introduces a complexity that does justify the small gain realized by the ACD CNO triggers Closing on the Down Transition of Any Signal that Could Have Opened the Window This is really just a variation of the previous method and, as such, inherits all its advantages and disadvantages. It is included only for completeness. The conclusion is the same, the gains do not warrant the additional complexity. The set of signals allowed to close the window is expanded to included any signal that could have opened a trigger window and that make an up transition during the window. This latter condition prevents signals initiated by previous events from closing this window. This pathology is illustrated by the following example. TKRS x represents the TKR 3-in-arow from different tower and TRGW is the trigger window signal. TKRS ACDS CALS Cal 1 Cal 2 TKRSx TKRSx TRGW Window 1 Window Figure 3 Timing Diagram Illustrating Erroneous Window Closure The first trigger window is a CAL only trigger, initiated by the first CALS pulse. The TKRS x pulse is either from the first CAL event or is from event occurring later in time. Since the trigger window is already active when TKRS x becomes active, TKRS x is not allowed to open a new trigger window. However, without the rule that only signals

9 Page 8 of 13 initiating the window or signals that make an up transition within a window are eligible to close the window, the trailing edge of TKRS x would prematurely close the second trigger window as illustrated by the dotted line in trigger window 2. In principle, this method does offer some improvement as illustrated in the following example. TKRS ACDS CALS TRGW Figure 4 Timing Diagram Illustrating Optimal WIndow Closure The CALS signal does make an up transition within the trigger window and, by definition, is allowed to close the window. The correct data is present within the trigger window and, therefore, the trigger window can be safely shorten by the amount indicated by the dotted lines. This technique improves the timing of the trigger acknowledge signal and shortens the deadtime associated with the trigger formation. However, given the parameters of the LAT trigger signals, the benefit is small. Potentially 10-20% could be shaved off the deadtime associated with the trigger window and result in a ~100 nsec less jitter on the trigger acknowledge signal. This is deemed not worth the effort Window Closes at a Fixed Time after Window Open This is the simplest to understand and likely the simplest to implement. The trigger window width is set at a fixed, but programmable width, meant to cover the longest jitter time. In practice, this time would likely be set at the TKR jitter time plus a little margin, something like 550nsec. The disadvantages are not closing the window as soon as possible and the burden of implementing the one shot. The benefits of closing the window as soon as possible have already been determined to be small. Implementing one programmable one-shot must be small potatoes on the scale of the GLT. This is the favored choice. 4.3 The Trigger Vector To form a trigger one needs the value of all the trigger signals during the time the trigger

10 Page 9 of 13 window is opened. The previous sections have glossed over precisely how this is done. There are two ways to do this. The first method involves stretching all signals to the maximum jitter. Since the stretched signals are up longer, there is an increased chance of values of a 1 being incorrectly latched. The greatest vulnerability is stretching the ACD signal, where false coincidences will lead to false vetoes, directly effecting gamma efficiency. The second method involves latching signals that have a value of 1 anytime the trigger window is opened. This includes not only signals that make a 0 to 1 transition, but signals that have may be 1 when the window opens and the signal(s) that opened the window. The values of the latched signals are captured at window close and form the basis of the trigger decision. The values are then cleared in anticipation of the next trigger window. This process is illustrated below. TKRS ACDS CALS TRGW TKRL ACDL CALL Figure 5 Latching the Trigger Vector The first three signals represent the usual TKRS, ACDS and CALS trigger signals. The TRGW signal illustrates the possible trigger window range, where the window may open anywhere within the light gray area and will close 500 nsec after it is opened, i.e. anywhere within the dark gray area. The dark lines represent typical signals. (This is the same notation used though-out this note, but it s a long note and one forgets.) The last three signals represent the latched values of their corresponding trigger signals. Note that the rising edge of the TKRS, ACDS and CALS trigger signals is always contained within the limits of the trigger window, and so, will be latched if that signal makes an up transition.

11 Page 10 of ACD Hit-Map Values and their Usage The values of the ACD signals captured at trigger window close can be used as the ACD veto hit-map, relieving the ACD FREE boards from this function. The simplification is the result of two conditions This function is a natural part of the trigger formation Only one set of ACD signals are under consideration at any given time. (Explained in greater detail below.) The first simplification is easy to understand. The trigger latches its input data as part of its process. The second is a bit more involved and explained below. The working decision up to this point has been to latch the ACD veto hit-map on the ACD FREE board. To do this, the state of these bits must be preserved until the trigger acknowledgement signal is received. In addition, because the trigger acknowledgement signal inherits all the jitter of the signals that generated it plus its own, the ACD veto hitmap signals must also be stretched to cover this uncertainty. In essence, the ACD FREE board must maintain a time history of the ACD signals. The stretching will also introduce the possibility of interpreting an ACD hit which should have been a 0 as a 1, leading to false vetoes. The process of latching the ACD within the trigger window naturally implements this stretching. This tactic dynamically and naturally minimizes the stretch. Having said all that, truth in advertising demands that it is said that a back-of-the-envelop calculation indicates that this is likely a small effect. The guess is that the false veto rate is the probability of overlapping a random ACD signal with a trigger window containing a real gamma. Assuming 10 tiles veto 1 tower and each tile runs at 1KHz, the ACD veto signal for each tower is active 10KHz * ACD width ( nsecs) =.1-.5%. Further assuming that a gamma event touches of the average of 2 towers, then the probability that gamma tower overlaps a random ACD signal is twice this, i.e %.) Latching the ACD signals at the trigger makes this time as short as possible favoring the.2%, stretching favors the 1% number. As an aside, this is why the ACD signal should be as short as possible. If the ACD signal is held high until the raw signal (the signal that is being discriminated) returns to baseline, this could be in the 1-2usec range, leading to 2-4% false veto rates. Note that this extension of the ACD signal is distinct from the stretching needed to ensure proper capture. This is elaborated on in the next section. The more relevant argument is that capturing on the GLT is simpler. It has a side benefit of (arguably) being a more faithful representation of the actual values Arguments For and Against Latching on the FREE Board/Extending ACD Signal The ACD design originally had the discriminated pulse being held until the raw signal returned to baseline. Until the signal returns to the baseline, the discriminator cannot refire. During this time, the ACD can provide no information. The safest thing to do, the argument went, is to assume the worst and hold the signal high, in some sense assuming that, yes

12 Page 11 of 13 indeed, a particle did pass during this ambiguous time. The argument on the other side is that, at 1KHz the probability of another particle passing is very small. The right thing to do is to assume the opposite, that there is no particle in the ambiguous time. This argument was countered by the ACD group with the true statement that small must be small compared to the requirements of 99.97% efficiency, i.e. small is <.03%. This argument is countered by, yes that is true, but this additional efficiency is only needed in the final analysis level, not at the trigger level. If the hardware level 1 trigger is used, the false veto rate is capped somewhere between the rate dictated by nsecs stretch times. However, if the VETO hit-map (commonly referred to as the throttle) is not used in the level 1 trigger, then this logic must be implemented in the software. If the ACD hit-map is formed from the signals that stay high until the level drops to baseline, the same problem of the false vetoes reappears, only this time at the software level. The ACD countered with the argument that the software has an advantage over the hardware level 1 trigger. It can point the track precisely at a particular tile, thus lowering the false veto rate by the typical number of tiles (4-10) that shadow a tower. Unfortunately, it is unlikely that the software can do track finding at 10KHz due to finite CPU resources, thus forcing the software to use only a slightly improved version of the hardware algorithm. This leaves only two choices Live with the false veto rate Let the few events that should have been vetoed sneak through the first stage of filtering and try to remove them by some other means. This note favors the latter strategy, not so much because of the loss in efficiency caused by the first strategy is too high, but because this loss cannot be measured or monitored in any reasonable way. In order to remove these events, the software would use tracking information to point the track at a particular ACD tile and then use the ACD pulse height information. Using tracking information is permitted because this processing is occurring only after the event rate has been reduced to < 100Hz. The pulse height information has the same properties as a stretched signal, i.e. it is a slow signal, so, it too could introduce a false veto. The false veto rate is reduced because only 1 tile is examined. This method also has the advantage that the data can be used to get a handle on the magnitude of this problem. The conclusion of this section is that the favored solution is to capture the VETO hit-map on the GLT. If this is deemed impossible, then the second choice is to capture the VETO hit-map on the FREE board, but to stretch the signals only by the amount needed to cover the trigger jitter. In both cases, the ACD pulse height information would be used to do the final cleanup. A conversation with the ACD group indicated that using the pulse height may not be a good idea, but they did not elaborate on exactly why. Steve Ritz apparently thinks it is okay since he originally mentioned the use of the pulse height information to the FSW GROUP. Need some resolution here.

13 Page 12 of Other Consequences of Capturing the VETO Hit-Map on the GLT As stated above, capturing the VETO hit-map on the GLT is the favored solution. Doing this may cause GLT implementation problems stemming from the large number of signals. The GLT must handle ~275 input signals (no choice here, also x2 since these are LVDS) 216 ACD VETO signals 12 ACD CNO signals 16 TKR 3-in-a-row signals 16 CAL LO signals 16 CAL HI signals In principle the output of the trigger can be as little as 2 serial streams, The trigger message The GLT s contribution to the data. Capturing the ACD VETO hit map on the GLT prohibits one from combining the ACD signals before GLT processing. It is the GLT that is producing the latched versions of these signals so all these signals must be input to the GLT. This presents a pin count problem if the GLT is implemented in the standard FPGA, having on the order of IO pins. Spreading the IO pins over 4 FPGAs is possible, but then the latched trigger information must be recombined and passed to another FPGA for the actual trigger formation. In the simplest scenario, the first stage of trigger processing produces a trigger strobe and the trigger vector. Essentially one needs 275 input signals and 275+ output signals. Other schemes have been explored to reduce the number of FPGAs from a maximum of 7 to 2-3. These involve multiplexing signals in at twice the nominal frequency or making some compromises in the allowable combinations. This is as far as this note wishes to go on this subject, leaving its final resolution to the electronics experts. The purpose of this section is only to note the problem, not to solve it. 5 Forming the Trigger This section assumes that a trigger window has been opened, then closed, and that the resulting trigger vector is available for examination. The trigger vector is first reduced to the trigger primitives by appropriately combining the signals. ACD Veto Hit Map produces o 16 Tower Veto Signals o 15 Signals, 3 from each of the 5 faces representing >0, >1, >2 hits on a face o 1 CNO Signal o 16 TKR 3-in-a-row

14 Page 13 of 13 o 1 CAL HI o 1 CAL LO LAT-SS-286, The Conceptual Design of the GLT, describes how these signals are used to produce the LAT triggers. (Given that this is merely a conceptual design, the implementer may choose a solution that maps the latched signals directly to the trigger message.) 6 Control Signals Each input trigger signal to the GLT must have two static configuration bits A standard enable/disable An window enable/disable The standard enable/disable is used to test the trigger and remove misbehaving inputs. The window enable/disable is used to indicate whether a particular signal can or cannot open/close a window. In practice, all but the ACD VETO signals would be enabled. The ACD signals are prohibited from opening a window since, in general, they only veto a trigger. The programmability is there for testing purposes.