20.01.2015 Views

Performance Modeling and Benchmarking of Event-Based ... - DVS

Performance Modeling and Benchmarking of Event-Based ... - DVS

Performance Modeling and Benchmarking of Event-Based ... - DVS

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

5.1. SPECJMS2007 - A STANDARD BENCHMARK 99<br />

Audit Test Scope Description<br />

Input rate is within +-5%<br />

<strong>of</strong> configured value<br />

Interaction For each Interaction, the observed input rate is calculated<br />

as count/time for the measurement period.<br />

This must be within 5% <strong>of</strong> the value prescribed by<br />

Total message count is<br />

within +-5% <strong>of</strong> configured<br />

value<br />

Input rate distribution deviations<br />

do not exceed<br />

20%<br />

90th percentile <strong>of</strong> Delivery<br />

Times under 5000ms<br />

All messages sent were received<br />

Interaction<br />

Interaction<br />

Message<br />

Destination<br />

Message<br />

Destination<br />

the topology.<br />

Using a model <strong>of</strong> the scenario, the benchmark knows<br />

how many messages should be processed as part <strong>of</strong><br />

each Interaction. The observed number <strong>of</strong> messages<br />

sent <strong>and</strong> received by all parties must be within 5% <strong>of</strong><br />

this value.<br />

The percentage <strong>of</strong> pacing misses (in driver threads)<br />

the benchmark will allow. A miss is when the timing<br />

code found itself behind where it thought it should<br />

be.<br />

Messages are timestamped when sent <strong>and</strong> received.<br />

The consequent delivery time is recorded as a histogram<br />

(for the measurement period only) in each<br />

event h<strong>and</strong>ler <strong>and</strong> the 90th percentile <strong>of</strong> this histogram<br />

must be on or under 5000ms.<br />

Fails if the final results (taken after the Drain Period)<br />

show that not all sent messages were received. For<br />

publish-subscribe topics this means received by all<br />

subscribers.<br />

Table 5.8: Audit Tests<br />

Auditor<br />

To validate the correctness <strong>of</strong> a benchmark run, SPECjms2007 comes with two auditor components:<br />

• Pre-Auditor (executed in Phase I): The pre-auditor checks the availability <strong>of</strong> JMS<br />

message destinations (queues / topcis) <strong>and</strong> whether messages exist before starting a benchmark<br />

run which would invalidate a run (but might otherwise only be detected when the<br />

test is completed). If the pre-audit fails for any reason the benchmark is not started.<br />

• Auditor (executed in Phase III): The auditor component is responsible for making<br />

sure that the run has been executed properly <strong>and</strong> that the results are valid. The Auditor<br />

is automatically launched at the end <strong>of</strong> the run to validate the results. It performs an<br />

explicit audit <strong>of</strong> the results against the specified requirements marking each audit test as<br />

P ASS or F AIL in the summary reports. A list <strong>of</strong> all audit tests in provided in Table 5.8.<br />

Minimizing the Impact <strong>of</strong> Non-MOM-Related Components<br />

As discussed in Section 5.1.1, the SPECjms2007 workload should be focused on measuring<br />

the performance <strong>and</strong> scalability <strong>of</strong> a MOM server’s s<strong>of</strong>tware <strong>and</strong> hardware components <strong>and</strong><br />

minimize the impact <strong>of</strong> other components that are not directly related to the MOM services.<br />

While developing SPECjms2007 two concerns had to be addressed in order to achieve this goal.<br />

The first one is how to avoid using a database <strong>and</strong> the second one is how to minimize the<br />

message processing overhead on the client. Given the fact that MOM servers, in their role as<br />

mediators in interactions, are typically less loaded than database servers or application servers,

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

Saved successfully!

Ooh no, something went wrong!