Information Technology Reference
In-Depth Information
Tabl e 1.
System Parameters
Tabl e 2.
Network Parameters
Network Configuration
Parameters
Component Parameters
No. of tiles 16
Processor UltraSPARCIII+
L1 I/D cache 64KB, 4-way
L2 cache bank 256KB 8-way
Memory bank 1GB, 4KB/page
Flit Size
16 bytes
Buffer Size
4
Pipeline Stage
5-stage
VCs per Virtual Network
4
Number of Virtual Networks
5
4.3 Simulation Results and Analysis
We compared our proposed method with baseline shared cache design (shared
tiled CMP using true LRU). We ran each benchmark up to its termination with
16 parallel threads. As mentioned in the documentation of PARSEC distribu-
tion, we only measured the performance of ROI (region of interest) section for
each benchmark. The performance is analyzed for two different configurations:
(a) 50% LP (b) 25% LP. Below we discuss the performance analysis of each
configuration separately.
50% LP.
In this configuration, we consider that the
LP
takes 50% ways (4
ways out of 8) from each set. Other important parameters are mentioned in
table 1. Figure 1 shows the performance comparison of our proposed method
with the baseline design. Each graph in the figure shows the results of different
performance metrics normalized to the baseline design value.
Figure 1(a) shows that random-LRU gets 21%-58% reduction in miss rate
with an average of 36.0%. Reduced off-chip trac due to lesser misses results
in lesser network usage. Hence network power consumption also reduces. Figure
1(b) shows the performance comparison in terms of cycle per instruction (CPI).
It shows that in case of random-LRU, CPI improves between 3.5%-47.8% with
an average improvement of 11.4%. Table 3 shows the performance improvement
in terms of both miss rate and CPI for each benchmark.
25% LP.
In this configuration, we consider that the
LP
takes 25% ways (2
ways out of 8) from each set. Figure 2 shows the performance comparison of our
proposed method with the baseline design.
Tabl e 3.
Performance improvement (in %) chart for Random-LRU (25% LP) with
baseline design
Benchmark Name (short name) CPI Miss Rate
fluidanimate (fluid)
3.54 %
49.81 %
x264 (x64)
47.80 % 39.61 %
ferret (fert)
14.24 % 21.59 %
freqmine (freq)
20.30 %
33 %
bodytack (body)
5.56 %
26.29 %
vips (vips)
8.20 %
58.76 %
Geometric mean
11.40 %
36 %
