From: Milian Wolff <mail@milianw.de>
Date: Mon, 7 Dec 2015 22:10:58 +0000 (+0100)
Subject: Optimize: Reduce memory cost for tracking active allocations.
X-Git-Tag: submit/tizen/20180620.112952^2~250
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=613cd167d80dd14cb5af2960b5a89e0ed60daf08;p=sdk%2Ftools%2Fheaptrack.git

Optimize: Reduce memory cost for tracking active allocations.

We split the hash into two, one for small allocations up to
numeric_limits<uint32_t>::max, and one for the bigger rest. This
allows us to reduce the cost by up to a factor of 2 for some apps.
A notable example here is heaptrack_gui itself, as it does tons of
small allocations for its lookup tables, which are long-lived.

In one test case this reduces the memory consumption of heaptrack by
up to 1GB for me.
---

diff --git a/accumulatedtracedata.cpp b/accumulatedtracedata.cpp
index 3440409..c734d93 100644
--- a/accumulatedtracedata.cpp
+++ b/accumulatedtracedata.cpp
@@ -55,7 +55,7 @@ AccumulatedTraceData::AccumulatedTraceData()
     traces.reserve(65536);
     strings.reserve(4096);
     allocations.reserve(16384);
-    activeAllocations.reserve(65536);
+    activeSmallAllocations.reserve(65536);
     stopIndices.reserve(4);
     opNewIpIndices.reserve(16);
 }
@@ -214,7 +214,13 @@ bool AccumulatedTraceData::read(istream& in)
                 continue;
             }
 
-            activeAllocations[ptr] = {traceId, size};
+            if (size <= std::numeric_limits<uint32_t>::max()) {
+                // save memory by storing this allocation in the list of small allocations
+                activeSmallAllocations[ptr] = {traceId, static_cast<uint32_t>(size)};
+            } else {
+                // these rare allocations consume more memory to track, but that's fine
+                activeBigAllocations[ptr] = {traceId, size};
+            }
 
             auto& allocation = findAllocation(traceId);
             allocation.leaked += size;
@@ -239,15 +245,11 @@ bool AccumulatedTraceData::read(istream& in)
                 cerr << "failed to parse line: " << reader.line() << endl;
                 continue;
             }
-            auto ip = activeAllocations.find(ptr);
-            if (ip == activeAllocations.end()) {
-                if (!fromAttached) {
-                    cerr << "unknown pointer in line: " << reader.line() << endl;
-                }
+            const auto info = takeActiveAllocation(ptr);
+            if (!info.traceIndex) {
+                // happens when we attached to a running application
                 continue;
             }
-            const auto info = ip->second;
-            activeAllocations.erase(ip);
 
             auto& allocation = findAllocation(info.traceIndex);
             if (!allocation.allocations || allocation.leaked < info.size) {
@@ -283,7 +285,8 @@ bool AccumulatedTraceData::read(istream& in)
     }
 
     /// these are leaks, but we now have the same data in \c allocations as well
-    activeAllocations.clear();
+    activeSmallAllocations.clear();
+    activeBigAllocations.clear();
 
     if (!reparsing) {
         totalTime = timeStamp + 1;
@@ -340,3 +343,22 @@ bool AccumulatedTraceData::isStopIndex(const StringIndex index) const
 {
     return find(stopIndices.begin(), stopIndices.end(), index) != stopIndices.end();
 }
+
+BigAllocationInfo AccumulatedTraceData::takeActiveAllocation(uint64_t ptr)
+{
+    auto small = activeSmallAllocations.find(ptr);
+    if (small != activeSmallAllocations.end()) {
+        auto ret = small->second;
+        activeSmallAllocations.erase(small);
+        return {ret.traceIndex, ret.size};
+    }
+    // rare
+    auto big = activeBigAllocations.find(ptr);
+    if (big != activeBigAllocations.end()) {
+        auto ret = big->second;
+        activeBigAllocations.erase(big);
+        return ret;
+    }
+    // happens esp. when we runtime-attached
+    return {};
+}
diff --git a/accumulatedtracedata.h b/accumulatedtracedata.h
index 5490a54..975d02c 100644
--- a/accumulatedtracedata.h
+++ b/accumulatedtracedata.h
@@ -125,9 +125,25 @@ struct MergedAllocation : public AllocationData
 };
 
 /**
- * Information for a single call to an allocation function
+ * Information for a single call to an allocation function for small allocations.
+ *
+ * The split between small and big allocations is done to save memory. Most of
+ * the time apps will only do small allocations, and tons of them. With this
+ * split we can reduce the memory footprint of the active allocation tracker
+ * below by a factor of 2. This is especially notable for apps that do tons
+ * of small allocations and don't free them. A notable example for such an
+ * application is heaptrack_print/heaptrack_gui itself!
+ */
+struct SmallAllocationInfo
+{
+    TraceIndex traceIndex;
+    uint32_t size;
+};
+
+/**
+ * Information for a single call to an allocation function for big allocations.
  */
-struct AllocationInfo
+struct BigAllocationInfo
 {
     TraceIndex traceIndex;
     uint64_t size;
@@ -174,9 +190,12 @@ struct AccumulatedTraceData
 
     bool isStopIndex(const StringIndex index) const;
 
+    BigAllocationInfo takeActiveAllocation(uint64_t ptr);
+
     // indices of functions that should stop the backtrace, e.g. main or static initialization
     std::vector<StringIndex> stopIndices;
-    std::unordered_map<uint64_t, AllocationInfo> activeAllocations;
+    std::unordered_map<uint32_t, SmallAllocationInfo> activeSmallAllocations;
+    std::unordered_map<uint64_t, BigAllocationInfo> activeBigAllocations;
     std::vector<InstructionPointer> instructionPointers;
     std::vector<TraceNode> traces;
     std::vector<std::string> strings;