1. Impact of parallelism on HEP software April 29 th 2013 Ecole Polytechnique/LLR Rene Brun

2. Software Upgrades All LHC experiments and groups like CERN/SFT are looking at all possible performance improvements or rethinking their software stack for the post LS2 years. This effort is driven by the new hardware and also the analysis of the hot spots. Work is going on in ROOT to support thread safety, parallel buffer merges and parallel Tree I/O. In the GEANT world, several projects(eg G4MT) investigate multi-core, gpus or like solutions. In this talk I will review the progress with one of these projects. 2

3. Hardware R.Brun : Paralllelism and HEP software 3 From a recent talk by Intel

4. If you trust Intel R.Brun : Paralllelism and HEP software 4

5. If you trust Intel 2 R.Brun : Paralllelism and HEP software 5

6. Vendors race R.Brun : Paralllelism and HEP software 6 parallelism

7. Parallelism: many failures R.Brun : Paralllelism and HEP software 7 inmos cray cm2 We failed in vectorizing codes like GEANT3 in 1985-1987 on CRAY, Cyber205, ETA10, IBM3090 because our approach was wrong Some successful attempts in online systems in 1983 We failed too on MPP systems like the Thinking Machines, Elxsi in 1991-1993 because our approach was wrong Are we going to take a wrong approach again?

8. R.Brun : Paralllelism and HEP software 8

9. Parallelism: key points R.Brun : Paralllelism and HEP software 9 Minimize the sequential/synchronization parts (Amdhal law): Very difficult Run the same code (processes) on all cores to optimize the memory use (code and read-only data sharing) Job-level is better than event-level parallelism for offline systems. Use the good-old principle of data locality to minimize the cache misses. Exploit the vector capabilities but be careful with the new/delete/gather/scatter problem Reorganize your code to reduce tails

10. Data Structures & parallelism R.Brun : Paralllelism and HEP software 10 event vertices tracks C++ pointers specific to a process C++ pointers specific to a process Copying the structure implies a relocation of all pointers I/O is a nightmare Update of the structure from a different thread implies a lock/mutex

11. Data Structures & Locality R.Brun : Paralllelism and HEP software 11 sparse data structures defeat the system memory caches Group object elements/collections such that the storage matches the traversal processes For example: group the cross- sections for all processes per material instead of all materials per process

12. Tools & Libs R.Brun : Paralllelism and HEP software 12 hbook zebra paw zbook hydra geant1 geant2 geant3 geant4 root minuit bos geant5

13. Detector Simulation tools 13 All based on the same principle: Sequential particle transport

14. The GEANT versions R.Brun : Paralllelism and HEP software 14 1975 1980 1990 1995 2010 G1 G2 G3 G4 functionality

15. Conventional Transport R.Brun : Paralllelism and HEP software 15 o o o o o o o o o o o o o o o o o o o o o o o o T1 T3 T2 o o o o o o o o o o o o o o o o o o o o T4 Each particle tracked step by step through hundreds of volumes when all hits for all tracks are in memory summable digits are computed

16. Analogy with car traffic R.Brun : Paralllelism and HEP software 16 2 5 3 1 4

17. Starting Assumptions The LHC experiments use extensively G4 as main simulation engine. They have invested in validation procedures. Any new project must be coherent with their framework. One of the reasons why the experiments develop their own fast MC solution is the fact that a full simulation is too slow for several physics analysis. These fast MCs are not in the G4 framework (different control, different geometries, etc), but becoming coherent with the experiments frameworks. Giving the amount of good work with the G4 physics, it is unthinkable to not capitalize on this work. R.Brun : Paralllelism and HEP software 17

18. Goals Design a new detector simulation tool derived from the Geant4 physics, but with a radically new transport engine supporting:  Full and Fast simulation (not exclusive)  Designed to exploit parallel hardware  this talk 18

19. Definitions 19 Detector Physical volumes Logical volumes ALICE4,354,7354,764 ATLAS29,046,9667,143 CMS1,166,3181,537 LHCb18,491,756709 A logical volume has a given shape and material

20. Steps/lvolume in Atlas R.Brun : Paralllelism and HEP software 20 Huge dynamic range 7100 lvolume types 29 million instances 7100 lvolume types 29 million instances

21. Simple observation: HEP transport is mostly local ! Locality not exploited by the classical transportation approach Existing code very inefficient (0.6- 0.8 IPC) Cache misses due to fragmented code Locality not exploited by the classical transportation approach Existing code very inefficient (0.6- 0.8 IPC) Cache misses due to fragmented code 50 per cent of the time spent in 50/7100 lvolumes 21

22. Neighbors/lvolume in Atlas R.Brun : Paralllelism and HEP software 22 Volumes with too many neighbors

23. Neighbors/lvolume in CMS R.Brun : Paralllelism and HEP software 23 Same problem with CMS

24. LHCB geometry statistics R.Brun : Paralllelism and HEP software 24 Better situation with neighbors because of a non cylindrical geometry 90 per cent of steps in 50/700 volumes

25. New Transport Scheme R.Brun : Paralllelism and HEP software 25 o o o o o o o o o o o o o o o o o o o o o o o o T1 T3 T2 o o o o o o o o o o o o o o o o o o o o T4 All particles in the same volume type are transported in parallel. Particles entering new volumes or generated are accumulated in the volume basket. All particles in the same volume type are transported in parallel. Particles entering new volumes or generated are accumulated in the volume basket. Events for which all hits are available are digitized in parallel

26. Tails again R.Brun : Paralllelism and HEP software 26 A killer if one has to wait the end of col(i) before processing col(i+1) Average number of objects in memory

27. A better solution R.Brun : Paralllelism and HEP software 27 Pipeline of objects Checkpoint Synchronization. Only 1 « gap » every N events Checkpoint Synchronization. Only 1 « gap » every N events This type of solution required anyhow for pile-up studies

28. A better better solution R.Brun : Paralllelism and HEP software 28 checkpoints At each checkpoint we have to keep the non finished objects/events. We can now digitize with parallelism on events, clear and reuse the slots. At each checkpoint we have to keep the non finished objects/events. We can now digitize with parallelism on events, clear and reuse the slots.

29. 29 Benchmarks/lessons from a prototype HT mode Excellent CPU usage Benchmarking 10+1 threads on a 12 core Xeon Locks and waits: some overhead due to transitions coming from exchanging baskets via concurrent queues Event re-injection will improve the speed-up 29

30. R.Brun : Paralllelism and HEP software 30

31. R.Brun : Paralllelism and HEP software 31

32. SFT S o F T w a r e D e v e l o p m e n t f o r E x p e r i m e n t s Vectorizing the geometry (ex1) R.Brun : Paralllelism and HEP software 32 Double_t TGeoPara::Safety(Double_t *point, Bool_t in) const { // computes the closest distance from given point to this shape. Double_t saf[3]; // distance from point to higher Z face saf[0] = fZ-TMath::Abs(point[2]); // Z Double_t yt = point[1]-fTyz*point[2]; saf[1] = fY-TMath::Abs(yt); // Y // cos of angle YZ Double_t cty = 1.0/TMath::Sqrt(1.0+fTyz*fTyz); Double_t xt = point[0]-fTxz*point[2]-fTxy*yt; saf[2] = fX-TMath::Abs(xt); // X // cos of angle XZ Double_t ctx = 1.0/TMath::Sqrt(1.0+fTxy*fTxy+fTxz*fTxz); saf[2] *= ctx; saf[1] *= cty; if (in) return saf[TMath::LocMin(3,saf)]; for (Int_t i=0; i<3; i++) saf[i]=-saf[i]; return saf[TMath::LocMax(3,saf)]; } Huge performance gain expected in this type of code where shape constants can be computed outside the loop

33. Vectorizing the geometry (ex2) R.Brun : Paralllelism and HEP software 33 G4double G4Cons::DistanceToIn( const G4ThreeVector& p, const G4ThreeVector& v ) const { G4double snxt = kInfinity ; // snxt = default return value const G4double dRmax = 100*std::min(fRmax1,fRmax2); static const G4double halfCarTolerance=kCarTolerance*0.5; static const G4double halfRadTolerance=kRadTolerance*0.5; G4double tanRMax,secRMax,rMaxAv,rMaxOAv ; // Data for cones G4double tanRMin,secRMin,rMinAv,rMinOAv ; G4double rout,rin ; G4double tolORMin,tolORMin2,tolIRMin,tolIRMin2 ; // `generous' radii squared G4double tolORMax2,tolIRMax,tolIRMax2 ; G4double tolODz,tolIDz ; G4double Dist,s,xi,yi,zi,ri=0.,risec,rhoi2,cosPsi ; // Intersection point vars G4double t1,t2,t3,b,c,d ; // Quadratic solver variables G4double nt1,nt2,nt3 ; G4double Comp ; G4ThreeVector Normal; // Cone Precalcs tanRMin = (fRmin2 - fRmin1)*0.5/fDz ; secRMin = std::sqrt(1.0 + tanRMin*tanRMin) ; rMinAv = (fRmin1 + fRmin2)*0.5 ; if (rMinAv > halfRadTolerance) { rMinOAv = rMinAv - halfRadTolerance ; } else { rMinOAv = 0.0 ; } tanRMax = (fRmax2 - fRmax1)*0.5/fDz ; secRMax = std::sqrt(1.0 + tanRMax*tanRMax) ; rMaxAv = (fRmax1 + fRmax2)*0.5 ; rMaxOAv = rMaxAv + halfRadTolerance ; // Intersection with z-surfaces tolIDz = fDz - halfCarTolerance ; tolODz = fDz + halfCarTolerance ; …… //here starts the real algorithm Huge performance gain expected in this type of code where shape constants can be computed outside the loop All these statements are independent of the particle !!!

34. SFT S o F T w a r e D e v e l o p m e n t f o r E x p e r i m e n t s Vectorizing the Physics This is going to be more difficult when extracting the physics classes from G4. However important gains are expected in the functions computing the distance to the next interaction point for each process. There is a diversity of interfaces and we have now sub-branches per particle type. R.Brun : Paralllelism and HEP software 34

35. Where are we now? Present status  Several investigations of possible alternatives for “extremely parallel – no lock” transport  Not much code written, several blackboards full  Some investigation on a simplified but fully vectorized model to prove vectorization gain  New design in preparation 35

36. Major points under discussion  How to minimize locks and maximize local handling of particles  How to handle hit and digit structures  How to preserve the history of the particles This point seems more difficult at the moment and it requires more design  What is the possible speedup obtained by micro- parallelisation  What are the bottlenecks and opportunities with parallel I/O 36

37. SFT S o F T w a r e D e v e l o p m e n t f o r E x p e r i m e n t s 37 Current design Input particle list Output particle list p array Hits p array History List of logical Volumes List of baskets for lv Active event list Sensitive volumes Digits for lv and event ev Logical Volume lv List of active events for lv Event ev Digitizer thread Events BF: basket status (one char per B) Transport thread Ev build thread Reused after each transport task Flushed at the end of event

38. Features  Pros Excellent potential locality Easy to introduce hits and digits  Cons One more copy (but it is done in parallel) More difficult to preserve particle history (it is non-local!) and introduce particle pruning 38

39. Processing flow I  The transport thread takes particles from the input buffer and transports them till they stop, interact or exit from the volume At this point they are inserted in the output particle buffer for further processing If the LV is a sensitive detector, hits are generated and stored per LV basket A LV basked history record is kept (under investigation)  Input and output particle buffers are fixed size structures, which can however evolve (be optimised) during simulation 39

40. 40 Design under study Input particle list Output particle list p array Hits p array History List of logical Volumes List of baskets for lv Active event list Sensitive volumes Digits for lv and event ev Logical Volume lv List of active events for lv Event ev ✔ full! ✗ empty! BF: basket status (one char per B)

41. Note Containers are “slow growing” contiguous containers Every time a container has to grow, it is realloc-ated contiguously to the new size  A blocking operation We expect containers size to converge  If not, there is a design problem 41

42. Processing flow II  When an input particle buffer is exhausted It is marked as such by the transport thread in the LV#BF (Logical Volume # Basket Flag)  Then the transport thread scans the LV#BF (Logical Volume # Basket Flag) data structure to find the next basket to be transported  Used buffers are scanned by the dispatcher thread that updates a global track counter per event  And then they are declared available to be filled (a) to be reused 42

43. Important!! a(available) basket being filled by the dispatcher f(full) basket ready to be dispatched r(ready) basket ready to be transported t(transporting) basket being transported 43

44. 44 Current design Input particle list Output particle list p array List of baskets for lv Logical Volumes tt1) The transport thread has finished working on the input array tt2) The transport thread marks the lv#bn from transporting (t) to “to be dispatched” (f) LV BN t dt1) The dispatch takes the first f basket and dispatches the output particles into the input particle lists of the baskets available to be filled (a) tt3) The transport thread gets a basket to be transported (r) from the fast selection list and marks it “transporting” (t) f a Dispatcher a dt2) When dispatching is finished the basket is moved from f to a a dt3) When a input list is full, the basket is moved from a to r, ready to be transported and it is pushed into the fast selection list tt0) Initial status: the transport thread is transporting a basket, marked as “transporting” (t) Input particle list Output particle list p array List of baskets for lv LV BN ra r ta Transport thread Dispatch thread BF: basket status asynch!! Fast LV#BN queue LV#BN

45. Locks…  The only lock is the push and pop from the fast selection queue  The dispatcher watches continuously the done byte-vector and dispatch every new basket that is ready Or it can sleep some and then process a number of done baskets in succession  The transport thread marks the done basket (no lock!) No one touches a (t) basket apart the transport thread that deals with it  The transport thread gets a new basket (lock!) This is to avoid that two threads get the same basket or that the dispatcher thread is updating the fast selection queue  The dispatcher thread does not need to lock the whole bit-array while dispatching The basket in f or in a will not be touched by the transport threads  The only doubtful situation is when there are no basket to be transported… In this case the “global” threshold for transporting should be lowered by a hungry transport thread (lock, but just to update an integer!) And the dispatcher will mark baskets as ready to be transported (r) 45

46. Memory We hope to have a self-adjusting system that will stabilise with time In case of an “accident” (an event much larger that any other), we need a way to “quench inflation” We have identified two methods  Event flushing: do NOT transport particles from a given set of events and move them directly to the output buffer  Energy flushing: transport low energy particles and move the high energy ones to the output buffer “Untransported” particles are just reinjected into the system, but they do not shower 46

47. Processing flow III  Note an important point The LV basket structure has input and output particle buffers and hits and history buffers  Input and output particle buffers are Multi-event Volatile, they get emptied and filled during transport of a single event  Hits and history buffers are Per event Permanent during the transport of a single event A basket of a LV can be handled by different threads successively, each one with a new input and output buffers …but all these threads will add to the Hits and history data structure till the event is flushed 47

48. Processing flow IV  When an event is finished, the digitizer thread kicks in and scans all the hits in all the baskets of all the LVs and digitise them, inserting them in the LV event->digit structure  When this is over, the event is built into the event structure (to be designed!) by the event builder thread  After that, the history for this event is assembled by the same thread If…  Then the event is output By an output thread or in parallel? 48

49. Questions?  How many dispatcher, digitizer and event-builder threads? Difficult to say, we need some more quantitative design work Measurements with G4 simulations could help  Transport thread numbers will have to adapt to the size of simulation and of the detector In ATLAS for instance 50% of the time is spent in 0.75% of the volumes Threads could be distributed proportionally to the time spent in the different LVs 49

50. 50 Short term tasks Continue the design work – essential before any more substantial implementation  This is the most important task at the moment  We have to evaluate the potential bottlenecks before starting the implementation Implement the new design and evaluate it against the first Demonstrate speedup of some chosen geometry routines  Both on x86 CPUs and GPUs Demonstrate speedup of some chosen physics methods  Particularly in the EM domain