MemPerf: Profiling Allocator-Induced Performance Slowdowns

In one paragraph

The memory allocator affects performance far beyond malloc and free: where it places objects decides cache and TLB behavior for every later access. No existing profiler could say that a slowdown comes from the allocator, so programmers optimize the wrong code. MemPerf identifies allocator-induced slowdowns without comparing against another allocator. It follows the whole life cycle of heap objects: a thread-aware, type-aware model estimates the cost of management operations, and a top-down method with hardware performance counters attributes slow memory accesses (cache misses, TLB misses, false sharing) to allocator decisions. It caught 98% of medium and large slowdowns with no false positives and exposed design issues in widely used allocators.

Key ideas

• Life-cycle based detection: an allocator can slow a program at allocation and deallocation time and at every access to the objects it placed, so both are measured separately.
• A thread-aware, type-aware performance model that predicts the slowdown of management operations against an oracle allocator, with no second run needed.
• Top-down attribution of slow memory accesses, using performance counters and lightweight sampling to measure allocator-induced cache and TLB misses correctly.
• Low interference: locks and relevant system calls are sampled at 1%, so the profiler is usable in deployment.

Results

• Reported 110 of 112 allocator-induced slowdowns above 5% (98.2%), with no false positives.
• About 7.5% profiling overhead (geometric mean), against 20% for perf, 56% for Feather, and 171% for CCProf in the same study.
• Pinpointed multiple known and previously unknown design issues in widely used allocators.