Semi-random initial value in the L1 trigger prescale counter #37506
Conversation
|
-code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-37506/29216
Code check has found code style and quality issues which could be resolved by applying following patch(s)
|
silviodonato
force-pushed
the
semiRandomInitL1PSCounter
branch
from
d38aed1
to
05f0658
Compare
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-37506/29218
|
|
A new Pull Request was created by @silviodonato (Silvio Donato) for master. It involves the following packages:
@epalencia, @cmsbuild, @cecilecaillol, @rekovic can you please review it and eventually sign? Thanks. cms-bot commands are listed here |
| @@ -164,6 +165,32 @@ namespace l1t { | |||
| /// pointer to Tau data list | |||
| inline const BXVector<const GlobalExtBlk*>* getCandL1External() const { return m_candL1External; } | |||
|
|
|||
| //initializer prescale counter using a semi-random value between [1, prescale value] | |||
| static const std::vector<double> semirandomNumber(const edm::Event& iEvent, | |||
There was a problem hiding this comment.
Why static, rather than just a const class member?
There was a problem hiding this comment.
I wanted to stress that this is a standalone function, that might be used even outside a GlobalBoard object.
I might define it as an external function eg. l1t::semirandomNumber(...), but I don't know in which file I should define this function. I've just learned that I cannot define it as l1t::GlobalBoard::semirandomNumber(...) because a namespace cannot have the same name of a class.
I don't have a strong opinion. What do you suggest?
There was a problem hiding this comment.
No strong opinion here, either. :)
I guess l1t::GlobalBoard::semirandomNumber(..) is effectively how this function would be called outside this class if using the current implementation. I don't see a good place for this elsewhere in this package, so unless L1T has a better idea, I would leave it here like this (maybe the implementation could be moved to the .cc file, though).
| } | ||
| return out; | ||
| } | ||
|
|
There was a problem hiding this comment.
Maybe this never happens for the values used by this function, but for doubles in general I think the asserts above could fail (given floating-point precision and the fact that out[i] is the result of a computation). Is that why they are there?
The current impl only handles the case of non-fractional prescales, right?
Should the randomisation be per algo (meaning, a different rand for each algo), or not (like in the current impl)?
In case it helps, I find the impl below a bit more readable (but I didn't test it, and there are likely better ways than using std::round to check for integers).
auto out = prescaleFactorsAlgoTrig;
// pick a random number from a combination of run, lumi, event numbers (different number for different threads)
std::srand(iEvent.id().run());
std::srand(std::rand() + iEvent.id().luminosityBlock());
// this causes different (semi)random number number for different streams
// reminder: different streams have different initial event number
std::srand(std::rand() + iEvent.id().event());
// (semi)random number in [0,1] range
auto const semirandom = std::rand() / double(RAND_MAX);
for (auto& ps : out) {
// if the ps is smaller than 1 (e.g. ps=0, ps=1) or if the prescale is fractional, it is not changed
// else, replace ps with a semirandom integer in the [1,ps] range
if (ps > 1 and ps == std::round(ps)) {
ps = std::round((ps - 1)*semirandom + 1);
}
}
return out;There was a problem hiding this comment.
It looks like my reply is lost.
-
I replaced the assert with a warning. In case of problems it will simply use the standard method ( take PS )
-
I think that it should work also with fractional prescales, as they don't have a specific treatment in GlobalBoard.cc (not 100% sure though)
-
randshould be common to among all the algos in order to preserve the synchronization of the L1 prescales (eg. L1_ZeroBias1 with PS = 20 should be always in the shadow of L1_ZeroBias2 with PS =10 ) -
thanks, I partially copied your code. I had to make some change to preserve the behavior described above.
| m_prescaleCounterAlgoTrig.push_back(semirandomNumber(iEvent, prescaleFactorsAlgoTrig)); | ||
| } else { | ||
| m_prescaleCounterAlgoTrig.push_back(prescaleFactorsAlgoTrig); | ||
| } |
There was a problem hiding this comment.
This looks like it could be done outside of the loop on totalBxInEvent (wouldn't we be recalculating the same semirandom PSs in each round of that loop, since run-lumi-event do not change?).
By the way, the m_prescaleCounterAlgoTrig.reserve call (above the diff) seems wrong. I was expecting m_prescaleCounterAlgoTrig.reserve(totalBxInEvent).
[..]
m_prescaleCounterAlgoTrig.reserve(totalBxInEvent);
auto const& prescaleCountersAlgoTrig = m_semiRandomInitialPSCounters ? semirandomNumber(iEvent, prescaleFactorsAlgoTrig) : prescaleFactorsAlgoTrig;
for (int iBxInEvent = 0; iBxInEvent <= totalBxInEvent; ++iBxInEvent) {
m_prescaleCounterAlgoTrig.push_back(prescaleCountersAlgoTrig);
[..]There was a problem hiding this comment.
Thanks I agree with both comments
silviodonato
force-pushed
the
semiRandomInitL1PSCounter
branch
from
05f0658
to
17833d5
Compare
| @@ -971,10 +971,13 @@ void l1t::GlobalBoard::runFDL(edm::Event& iEvent, | |||
| // prescale counters are reset at the beginning of the luminosity segment | |||
| if (m_firstEv) { | |||
| // prescale counters: numberPhysTriggers counters per bunch cross | |||
| m_prescaleCounterAlgoTrig.reserve(numberPhysTriggers * totalBxInEvent); | |||
| m_prescaleCounterAlgoTrig.reserve(totalBxInEvent); | |||
There was a problem hiding this comment.
@cms-sw/l1-l2 please note this is the only line which should affect the default behavior (semiRandomInitialPSCounters = false)
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-37506/29637
|
|
Pull request #37506 was updated. @epalencia, @cmsbuild, @cecilecaillol, @rekovic can you please check and sign again. |
| if (ps > 1) { | ||
| auto nps = semirandom - floor(semirandom / ps) * ps; | ||
| // if nps=0 or a wrong value (<0,>ps) use PS value (standard method) | ||
| if (nps > 0 || nps <= ps) |
There was a problem hiding this comment.
| if (nps > 0 || nps <= ps) | |
| if (nps > 0 and nps <= ps) |
No?
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-37506/30038
|
|
Pull request #37506 was updated. @epalencia, @cmsbuild, @cecilecaillol, @rekovic can you please check and sign again. |
|
please test I suggest to start testing. |
|
please test |
|
+1 Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-265261/24762/summary.html Comparison Summary@slava77 comparisons for the following workflows were not done due to missing matrix map:
Summary:
|
|
@cms-sw/l1-l2 do you have any comments? |
|
+l1 |
cmsbuild
added
fully-signed
l1-approved
and removed
l1-pending
pending-signatures
labels
|
This pull request is fully signed and it will be integrated in one of the next master IBs (tests are also fine). This pull request will now be reviewed by the release team before it's merged. @perrotta, @dpiparo, @qliphy (and backports should be raised in the release meeting by the corresponding L2) |
|
please test |
|
+1 Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-265261/24965/summary.html Comparison SummarySummary:
|
|
+1 |
PR description:
The standard behavior of the L1 trigger prescaler is to start a counter from the prescale value (eg. 1000) and then reduce it of one unit every time the corresponding L1 trigger is accepted. Once the counter reach 0, the L1 trigger is consider to pass also the L1 prescale, and the counter goes back to the prescale value.
The counter is reset at the beginning of each lumisection (unless you enable #37395)
This PR adds the option to set the initial prescale value of the PS counter from PS to a semirandom value taken in the range [0, PS]. This option is required to avoid that running on a heavily prescaled sample (eg. ZeroBias) the PS counter never reach 0 and therefore no events pass the prescaler. Starting on a random value, the average rate will be closer to the actual rate that we would see in the actual datataking (at P5 the L1 trigger evaluate a huge number of events per LS).
The "semi-random" refers to the
std::srandfunction which gives a reproducible sequence of random numbers.However the reproducibility is not preserved when running with multithreading because I use also the event number as input of the random number generator.
The current L1 trigger precaler does not preserve the reproducibility when multi-threading is on.
This PR is somehow related to #37395.
PR validation:
Running the prescaler on 3151 events (two LS: 1584 events + 1567 events) on ZeroBias with a prescale of 100 using 16 threads.
Default configuration: 8 events passed
Including this PR (default): 6 events passed
Including this PR (semiRandomInitialPSCounters on): 30 events passed
Second test (same configurations)
Default configuration: 4 events passed
Including this PR (default): 5 events passed
Including this PR (semiRandomInitialPSCounters on): 31 events passed
if this PR is a backport please specify the original PR and why you need to backport that PR:
Yes, I'd like to have a backport to 12_3_X - although it is not strictly necessary - as it will be used for the first stable beam and therefore for the first HLT menu.
We use this PR for the HLT rate estimate based on the L1 skim of the ZeroBias dataset.