Improving relocation performance in ZGC by identifying the size of small objects

Abstract: Modern Garbage Collectors provide performance improvements by increasing program locality to utilize the faster CPU cache. A common approach is to move objects together according to the mutators’ access order, which brings more relocations during GC. In most cases, more relocations would not impact performance when using concurrent Garbage Collectors such as ZGC. However, in constrained environments with fewer CPU cores or less memory, bad relocation performance will cause overall performance degradation. In this thesis, we investigated why larger objects do not benefit from better program locality, then proposed a new design to reduce the number of relocations by efficiently identifying and ignoring larger objects. As a result, the relocation performance can be improved. In constrained environments, this can lead to an increase in overall throughput. In the new design, we introduce an extra page type, the tiny page. If an object is considerably small that it could benefit from relocation, it will be placed on the tiny page when allocating. As a result, we could replace the time‐consuming size check of objects with a faster page type check. Memory fragmentation also can be reduced by this design. To evaluate this design, we add the size identification procedure into a locality improvement implementation named HCSGC. The results of benchmarks show a slight improvement in constrained environments. In the JGraphT benchmark, we see a 3‐5% speedup in different configurations with memory limitations. In the SPECjbb2015 benchmark, we see a ~1% increase in performance on average, but with overlapping confidence intervals. In the DaCapo benchmark suite, we see a 1% improvement in the sunflow benchmark with CPU constraint. For other benchmarks in DaCapo, no significant difference is discovered. The results suggest that the proposed new design is a feasible way of filtering out larger objects, and doing so can further improve the relocation and overall performance.