Fig 1.
Outline of the methodology.
Table 1.
System specifications.
Table 2.
Benchmarks tested, divided between integer (int column) and floating point (fp column).
Filled cells in columns “2006” and “2017” mean the benchmark appears in the corresponding suite. The columns labeled “_r” and “_s” refer to the application versions producing SPECrate and SPECspeed metrics, respectively. Columns “Inst.” and “#” show instruction count (x1012) and input identifier, respectively.
Fig 2.
MPKI1, MPKI2 and MPKI3 for all SPEC CPU2006 benchmarks, sorted by benchmark number.
Fig 3.
MPKI1, MPKI2 and MPKI3 for all SPEC CPU2017 single-threaded benchmarks, sorted by benchmark number.
Table 3.
Selected benchmarks and their performance metrics for minimum LLC size and no prefetching.
Fig 4.
MPKI3 vs. LLC size for the selected CPU2006 benchmarks, with and without prefetching.
Fig 5.
MPKI3 vs. LLC size for the selected CPU2017 benchmarks, with and without prefetching.
Fig 6.
Speed-ups enabled either by hardware prefetching, with the minimum cache size (X axis) or maximum LLC size, without prefetching (Y axis) over a baseline configuration without prefetching and minimum LLC size for the selected SPEC CPU2006 and CPU2017 benchmarks.
Integer and floating point benchmarks are represented by gray circles and black squares, respectively.
Fig 7.
CPI vs. MPKI3 for the selected CPU2006 benchmarks, varying LLC size and with prefetching (square marks) and without prefetching (x marks).
Slope units are cycles/miss. Slopes are comparable in all graphs because the ratio between X and Y scales is constant (10:1).
Fig 8.
CPI vs. MPKI3 for the selected CPU2017 benchmarks, varying LLC size and with prefetching (square marks) and without prefetching (x marks).
Slope units are cycles/miss. Slopes are comparable in all graphs because the ratio between X and Y scales is constant (10:1).
Fig 9.
Impact of the different hardware prefetchers on performance (CPI, bars) and bandwidth consumption (BPKI, line) for the selected CPU2006 benchmarks.
Fig 10.
Impact of the different hardware prefetchers on performance (CPI, bars) and bandwidth consumption (BPKI, line) for the selected CPU2017 benchmarks.
Fig 11.
Temporal evolution of MPKI3 and SimPoint selection for the selected CPU2006 benchmarks, with minimum LLC size and hardware prefetching.
Fig 12.
Temporal evolution of MPKI3 and SimPoint selection for the selected CPU2017 benchmarks, with minimum LLC size and hardware prefetching.