 |  |  | Effect of GC on Cache Performance |
Because the garbage collector moves data and code, it can significantly affect the memory system behavior of the code that subsequently executes. To investigate this effect, we used Performance Explorer to extract miss rates for the first (1), the first five (5), the first ten (10), and the first fifty (50) warehouse trace records immediately after a GC. Figures * and * present this data for instructions and data, respectively. As a reference, the All bars present the overall cache miss rate of the warehouse thread.
Instruction miss rates after a GC
From Figures * and * we see that both the instructions and data suffer increased cache misses immediately after a garbage collection. The misses in the L1 and L2 caches stabilize in less than 5 trace records. The much larger L3 cache takes even longer than 50 trace records to stabilize. The percentage of all references that are made to L2 is less than 0.30% for data and less than 0.05% for instructions for all but the first trace record after a GC. The percentage of all references that are to L3 is less than L2. The increased miss rates after a GC are also reflected in the IPC: the IPC in the first and first five trace records immediately following a GC are 0.069 and 0.357. In contrast the stable state IPC is 0.429.
These results clearly indicate that GC significantly degrades the memory performance of the application and it can take a significant amount of time, as indicated by the number of trace records, after a GC before the working set of the application is in the caches again. In future work we will investigate how other GC algorithms behave in this regard.
| | | Effect of GC on Cache Performance |