Check out the new USENIX Web site.
Memory Wall IssuesExperimentsExplorationGeneral Performance Trends

General Performance Trends

A graph over time of the warehouse thread's instruction per cycle and flushes per instructions completed

A graph over time of the warehouse thread's instruction per cycle and flushes per instructions completed.
 
AOS
Metric JIT-base start (Delta%) end (Delta%) JIT-opt0 (Delta%)
IPC 0.348 0.415 (+19.3%) 0.489 (+42.5%) 0.500 (+43.7%)
LSU_FLUSH/INST_CMPL 15.2% 7.0% (-54.2%) 0.3% (-97.4%) 0.1% (-99.3%)
INST_CMPL 4.18M 4.05M (+16.0%) 5.84M (+42.6%) 5.96M (+42.6%)
Metrics across JIT and AOS configurations.
 

The top graph in Figure * presents a graph of the IPC of a warehouse thread over time. For clarity of presentation, only the IPC trace records of the warehouse thread are displayed; all other Java threads are not shown. The noticeable gaps in the warehouse trace records are stop-the-world garbage collections, which disable Java thread scheduling when it runs.

Two general performance anomalies are noticeable in this graph. First, IPC improves over time. The IPC in the left corner is around 0.41, while the IPC in the right corner is above 0.49, a 20% increase. Second, before each of the GCs there is a significant IPC degradation or drop. We used the Performance Explorer to help us explore both of these phenomena.

  • Anomaly 1
  • Anomaly 2

  • Memory Wall IssuesExperimentsExplorationGeneral Performance Trends