Merge pull request #46848 from Dr15Jones/moduleEventAlloc

Added ModuleEventAllocMonitor Service

Author: cmsbuild

PerfTools/AllocMonitor/README.md

@@ -85,8 +85,8 @@ This service is multi-thread safe. Note that when run multi-threaded the maximum
 This service registers a monitor when the service is created (after python parsing is finished but before any modules
 have been loaded into cmsRun) and prints its accumulated information to the specified file at specified intervals. Both
 the file name and  interval are specified by setting parameters of the service in the configuration. The parameters are
-- filename: name of file to which to write reports
-- millisecondsPerMeasurement: number of milliseconds to wait between making each report
+- `filename`: name of file to which to write reports
+- `millisecondsPerMeasurement`: number of milliseconds to wait between making each report
 
 The output file contains the following information on each line
 - The time, in milliseconds, since the service was created
@@ -108,10 +108,27 @@ This service registers a monitor when the service is created (after python parsi
 have been loaded into cmsRun) and writes module related information to the specified file. The file name, an optional
 list of module names, and  an optional number of initial events to skip are specified by setting parameters of the
 service in the configuration. The parameters are
-- filename: name of file to which to write reports
-- moduleNames: list of modules which should have their information added to the file. An empty list specifies all modules should be included.
-- nEventsToSkip: the number of initial events that must be processed before reporting happens.
+- `filename`: name of file to which to write reports
+- `moduleNames`: list of modules which should have their information added to the file. An empty list specifies all modules should be included.
+- `nEventsToSkip`: the number of initial events that must be processed before reporting happens.
 
 The beginning of the file contains a description of the structure and contents of the file.
 
-This service is multi-thread safe.
+This service is multi-thread safe.
+
+
+### ModuleEventAllocMonitor
+This service registers a monitor when the service is created (after python parsing is finished but before any modules
+have been loaded into cmsRun) and writes event based module related information to the specified file. The service
+keeps track of the address of each allocation requested during an event from each module and pairs them with any
+deallocation using the same address. The list of addresses are kept until the event data products are deleted at which
+time the dallocations are paired with allocations done in a module. The list of addresses are then cleared (to keep memory usage down) but the amount of unassociated deallocations for each module is recorded per event.
+The file name, an optional list of module names, and  an optional number of initial events to skip are specified by setting parameters of the
+service in the configuration. The parameters are
+- `filename`: name of file to which to write reports
+- `moduleNames`: list of modules which should have their information added to the file. An empty list specifies all modules should be included.
+- `nEventsToSkip`: the number of initial events that must be processed before reporting happens.
+
+The beginning of the file contains a description of the structure and contents of the file. The file can be analyzed with the helper script `edmModuleEventAllocMonitorAnalyze.py`. The script can be used to find modules where the memory is being retained between Events as well as modules where the memory appears to be growing Event to Event. Use `--help` with the script for a full description.
+
+This service is multi-thread safe.

PerfTools/AllocMonitor/plugins/ModuleAllocMonitor.cc

@@ -32,106 +32,10 @@
 
 #include "moduleAlloc_setupFile.h"
 #include "ThreadAllocInfo.h"
+#include "ThreadTracker.h"
 
 namespace {
-  inline auto thread_id() {
-#if defined(ALLOC_USE_PTHREADS)
-    /*NOTE: if use pthread_self, the values returned by linux had
-     lots of hash collisions when using a simple % hash. Worked
-     better if first divided value by 0x700 and then did %. 
-     [test done on el8]
- */
-    return pthread_self();
-#else
-    return syscall(SYS_gettid);
-#endif
-  }
-
-  struct ThreadTracker {
-    static constexpr unsigned int kHashedEntries = 128;
-    static constexpr unsigned int kExtraEntries = 128;
-    static constexpr unsigned int kTotalEntries = kHashedEntries + kExtraEntries;
-    using entry_type = decltype(thread_id());
-    static constexpr entry_type kUnusedEntry = ~entry_type(0);
-    std::array<std::atomic<entry_type>, kHashedEntries> hashed_threads_;
-    std::array<std::atomic<entry_type>, kExtraEntries> extra_threads_;
-
-    ThreadTracker() {
-      //put a value which will not match the % used when looking up the entry
-      entry_type entry = 0;
-      for (auto& v : hashed_threads_) {
-        v = ++entry;
-      }
-      //assume kUsedEntry is not a valid thread-id
-      for (auto& v : extra_threads_) {
-        v = kUnusedEntry;
-      }
-    }
-
-    std::size_t thread_index() {
-      auto id = thread_id();
-      auto index = thread_index_guess(id);
-      auto used_id = hashed_threads_[index].load();
-
-      if (id == used_id) {
-        return index;
-      }
-      //try to be first thread to grab the index
-      auto expected = entry_type(index + 1);
-      if (used_id == expected) {
-        if (hashed_threads_[index].compare_exchange_strong(expected, id)) {
-          return index;
-        } else {
-          //another thread just beat us so have to go to non-hash storage
-          return find_new_index(id);
-        }
-      }
-      //search in non-hash storage
-      return find_index(id);
-    }
-
-  private:
-    std::size_t thread_index_guess(entry_type id) const {
-#if defined(ALLOC_USE_PTHREADS)
-      return (id / 0x700) % kHashedEntries;
-#else
-      return id % kHashedEntries;
-#endif
-    }
-
-    std::size_t find_new_index(entry_type id) {
-      std::size_t index = 0;
-      for (auto& v : extra_threads_) {
-        entry_type expected = kUnusedEntry;
-        if (v == expected) {
-          if (v.compare_exchange_strong(expected, id)) {
-            return index + kHashedEntries;
-          }
-        }
-        ++index;
-      }
-      //failed to find an open entry
-      abort();
-      return 0;
-    }
-
-    std::size_t find_index(entry_type id) {
-      std::size_t index = 0;
-      for (auto const& v : extra_threads_) {
-        if (v == id) {
-          return index + kHashedEntries;
-        }
-        ++index;
-      }
-      return find_new_index(id);
-    }
-  };
-
-  static ThreadTracker& getTracker() {
-    static ThreadTracker s_tracker;
-    return s_tracker;
-  }
-
+  using namespace cms::perftools::allocMon;
   using namespace edm::service::moduleAlloc;
   class MonitorAdaptor : public cms::perftools::AllocMonitorBase {
   public:
@@ -149,7 +53,7 @@ namespace {
   private:
     static ThreadAllocInfo& threadAllocInfo() {
       static ThreadAllocInfo s_info[ThreadTracker::kTotalEntries];
-      return s_info[getTracker().thread_index()];
+      return s_info[ThreadTracker::instance().thread_index()];
     }
     void allocCalled(size_t iRequested, size_t iActual, void const*) final {
       auto& allocInfo = threadAllocInfo();