Performance Modeling and Benchmarking of Event-Based ... - DVS
Performance Modeling and Benchmarking of Event-Based ... - DVS
Performance Modeling and Benchmarking of Event-Based ... - DVS
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
38 CHAPTER 4. PERFORMANCE ENGINEERING OF EVENT-BASED SYSTEMS<br />
This obvious relationship follows from the Utilization Law [62]. Once the CPU service times<br />
<strong>of</strong> events at the considered system node on the reference hardware configuration have been estimated,<br />
they can be used as reference values to extrapolate the service times to the node’s configuration<br />
in the target environment. Benchmark results (e.g., SPECcpu2006 or SPECjms2007)<br />
for the respective hardware configurations can be exploited in such extrapolations. If the chosen<br />
reference configuration is similar to the target configuration, no extrapolation is necessary. If,<br />
in addition to the CPUs, further resources at the system node are used when delivering events<br />
(e.g., secondary storage devices), the mean service times at these resources can be estimated<br />
based on the measured resource utilization in exactly the same way as shown above for the<br />
CPUs. In certain cases, techniques can be employed that help to estimate the service times<br />
without the need to do any measurements on the system [147]. Such techniques are based on<br />
analyzing how the system components are implemented at the code level.<br />
The mean network service time St,q<br />
NET can be calculated based on the available b<strong>and</strong>width <strong>and</strong><br />
the size <strong>of</strong> the message that has to be transferred (taking protocol overhead into account) when<br />
an event <strong>of</strong> type e t is sent over the network corresponding to connection c q : St,q NET = Mq/B t q .<br />
4.1.4 System Operational Analysis<br />
If we look at the CPUs <strong>of</strong> the nodes in our system as M/M/1 queues, we can use the following<br />
relationship from queueing theory to obtain an approximation for the mean response time Rt,j<br />
CP U<br />
<strong>of</strong> events <strong>of</strong> type e t at the CPU <strong>of</strong> node n j :<br />
From the Utilization Law it follows that:<br />
t=1<br />
Rt,j CP U = SCP t,j<br />
U<br />
1 − U CP U<br />
⎛<br />
|E|<br />
|E| |P |<br />
∑<br />
∑ ∑<br />
Uj CP U = λ t jSt,j CP U = ⎝<br />
t=1<br />
j<br />
k=1<br />
λ t,k<br />
j<br />
⎞<br />
(4.4)<br />
⎠ S CP U<br />
t,j (4.5)<br />
Similarly, it follows that the disk I/O utilization <strong>of</strong> node n j can be computed as<br />
⎛<br />
|E| |P |<br />
∑ ∑<br />
U I/O<br />
j = ⎝<br />
t=1<br />
k=1<br />
λ t,k<br />
j<br />
⎞<br />
⎠ S I/O<br />
t,j (4.6)<br />
<strong>and</strong> an approximation for the mean response time R I/O<br />
t,j<br />
subsystem <strong>of</strong> node n j is given by:<br />
R I/O<br />
t,j = SI/O t,j<br />
1 − U I/O<br />
j<br />
<strong>of</strong> events <strong>of</strong> type e t at the disk I/O<br />
(4.7)<br />
Analogously, an approximation for the mean response time R I/O<br />
t,j <strong>of</strong> events <strong>of</strong> type e t at<br />
the disk I/O subsystem <strong>of</strong> node n j can be obtained. The arrival rates λ t,k<br />
j can be derived by<br />
solving the system <strong>of</strong> simultaneous equations (4.1). Using the same approach, we can estimate<br />
the utilization <strong>of</strong> the network links in the system <strong>and</strong> the network response times. The rate at<br />
which events <strong>of</strong> type e t (published by any publisher) are sent over connection c q can be computed<br />
as follows:<br />
|P |<br />
∑ (<br />
τq t =<br />
k=1<br />
λ t,k<br />
l<br />
)<br />
ν t,k<br />
l,r<br />
|P |<br />
∑ (<br />
+<br />
k=1<br />
λ t,k<br />
r<br />
)<br />
ν t,k<br />
r,l<br />
(4.8)