Thread-Level Parallelism - Computer Architecture: A Quantitative Approach - page 403

FIGURE 5.29 The performance on the SPECjbb2005 benchmark for three multicore pro-
cessors as the number of processor chips is increased . Notice for this parallel bench-
mark, nearly linear speedup is achieved.
Performance And Energy Efficiency Of The Intel Core I7
Multicore
In this section, we examine the performance of the i7 on the same two groups of benchmarks
we considered in Chapter 3 : the parallel Java benchmarks and the parallel PARSEC bench-
marks (described in detail in Figure 3.34 on page 231). First, we look at the multicore perform-
ance and scaling versus a single-core without the use of SMT. Then, we combine both the mul-
ticore and SMT capability. All the data in this section, like that in the earlier i7 SMT evaluation
(Chapter 3, Section 3.13 ) come from Esmaeilzadeh et al. [2011] . The dataset is the same as that
used earlier (see Figure 3.34 on page 231), except that the Java benchmarks tradebeans and pjb-
b2005 are removed (leaving only the five scalable Java benchmarks); tradebeans and pjbb2005
never achieve speedup above 1.55 even with four cores and a total of eight threads, and thus
are not appropriate for evaluating more cores.
Figure 5.30 plots both the speedup and energy efficiency of the Java and PARSEC bench-
marks without the use of SMT. Showing energy eiciency means we are ploting the ratio of
the energy consumed by the two- or four-core run by the energy consumed by the single-core
run; thus, higher energy eiciency is beter, with a value of 1.0 being the break-even point. The
unused cores in all cases were in deep sleep mode, which minimized their power consump-
tion by essentially turning them of. In comparing the data for the single-core and multicore
benchmarks, it is important to remember that the full energy cost of the L3 cache and memory
interface is paid in the single-core (as well as the multicore) case. This fact increases the likeli-
hood that energy consumption will improve for applications that scale reasonably well. Har-
monic mean is used to summarize results with the implication described in the caption.