Parallel Computer Architecture

642 CHAPTER 8 Directory-Based Cache Coherence

Table 8.3 Characteristics of Microbenchmarks and Workloads Running on an Eight-

Quad NUMA-Q

Workload

Remote Read

Misses

Remote Write

Misses

TPC-B-like

TPC-D (Q9)

Latency of L

2

Misses

All

8,020 ns

9,350 ns

630 ns

580 ns

Remotely

Satisfied

8,300 ns

9,625 ns

4,300 ns

3,950 ns

SCLIC

Utilization

95%

95%

54%

40%

Percentage of L

2

Misses Satisfied in

Local

Memory

1.5%

1%

80%

85%

Other Local Local

Cache "Remote

Cache"

0% 2%

0% 2%

2%

11.5%

5.5% 4%

Remote

Node

96.5%

97%

6.5%

5.5%

of about 10 to 1. However, the inclusion of a remote access cache keeps the fre-

quency of artifactual communication very low. Th