01.12.2012 Views

Architecture of Computing Systems (Lecture Notes in Computer ...

Architecture of Computing Systems (Lecture Notes in Computer ...

Architecture of Computing Systems (Lecture Notes in Computer ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

56 J. Zeppenfeld and A. Herkersdorf<br />

3 Simulation Results<br />

Given the network processor MP-SoC presented above, this section presents simulation<br />

results that compare a classical system implementation with one enhanced by an<br />

autonomic layer. In order to show the adaptive nature <strong>of</strong> the autonomic system, various<br />

types <strong>of</strong> packet bursts, each consist<strong>in</strong>g <strong>of</strong> 1000 packets, are sent <strong>in</strong> alternat<strong>in</strong>g<br />

fashion to the <strong>in</strong>put MAC for process<strong>in</strong>g. The packet <strong>in</strong>ter-arrival rate is fixed at 4 μs,<br />

result<strong>in</strong>g <strong>in</strong> a burst duration <strong>of</strong> 4 ms <strong>in</strong>dependent <strong>of</strong> the packet process<strong>in</strong>g duration.<br />

For each <strong>of</strong> the <strong>in</strong>com<strong>in</strong>g burst scenarios, the system must either adapt to meet the<br />

process<strong>in</strong>g requirements, or ma<strong>in</strong>ta<strong>in</strong> a sufficient reserve <strong>of</strong> process<strong>in</strong>g power to be<br />

able to cope with even the largest and most process<strong>in</strong>g <strong>in</strong>tensive packets.<br />

All systems used below are based on a common base system configuration. The<br />

hardware components and their functionality, as well as the s<strong>of</strong>tware tasks used for<br />

packet process<strong>in</strong>g were described <strong>in</strong> sections 2.1 and 2.2, respectively. Unless stated<br />

otherwise, the <strong>in</strong>itial task distribution is such that one CPU is responsible for <strong>in</strong>gress<br />

path process<strong>in</strong>g (Tasks 1 and 2), one CPU is responsible for payload process<strong>in</strong>g (Task<br />

3), and the third CPU is responsible for egress path process<strong>in</strong>g (Tasks 4 and 5). The<br />

<strong>in</strong>itial CPU frequencies are chosen such that the system would be able to handle all<br />

burst scenarios without requir<strong>in</strong>g parameter adaptation. Those systems conta<strong>in</strong><strong>in</strong>g<br />

autonomic enhancements use the monitors, actuators and LCT evaluator presented <strong>in</strong><br />

section 2.3.<br />

3.1 Comparison <strong>of</strong> Autonomic and Static <strong>Systems</strong><br />

The simulation results shown <strong>in</strong> Figure 5 compare the objective value and frequency<br />

adjustment <strong>of</strong> four different system configurations. The first two systems are static<br />

without any autonomic enhancements, where the first corresponds to the base system<br />

configuration described above. In the second static system, both the task distribution<br />

and CPU frequency were hand optimized with prior knowledge <strong>of</strong> the <strong>in</strong>com<strong>in</strong>g<br />

packet traffic. The results <strong>of</strong> this optimized system demonstrate how an ideally parameterized<br />

static system compares to a system with autonomic enhancements.<br />

The two autonomic systems are both based on the common system configuration,<br />

but are differentiated by the monitor signals available to them. While the second system<br />

uses the AE <strong>in</strong>terconnect to share workload <strong>in</strong>formation globally among the CPU<br />

AEs, the first autonomic system must base its optimization decisions solely on local<br />

monitor <strong>in</strong>formation. Both systems rema<strong>in</strong> capable <strong>of</strong> migrat<strong>in</strong>g tasks, however.<br />

In comparison to a static system, the trend <strong>of</strong> the objective value clearly shows the<br />

benefits <strong>of</strong> an autonomic system regard<strong>in</strong>g fulfillment <strong>of</strong> the objective function chosen<br />

<strong>in</strong> section 2.3.3. The autonomic system with global <strong>in</strong>formation is able to ma<strong>in</strong>ta<strong>in</strong><br />

system operation with<strong>in</strong> approximately 10% <strong>of</strong> the optimum for all <strong>in</strong>com<strong>in</strong>g<br />

traffic scenarios. Although the hand optimized static system is able to top this for the<br />

fifth and most process<strong>in</strong>g <strong>in</strong>tensive traffic scenario, it is not able to ma<strong>in</strong>ta<strong>in</strong> such a<br />

high level <strong>of</strong> optimization across all bursts. This is a direct consequence <strong>of</strong> the fact<br />

that a system optimized for one scenario is not necessarily optimized for other scenarios.<br />

Whereas the designer must choose a certa<strong>in</strong> scenario for which the static system

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!