Posted by admin | Posted on 16-03-2012
Category : Computer
Tags: computer measuring
Perturbations due to measuring
One of the curious (and certainly most annoying!) aspects of developing tools to measure computer-systems performance is that instrumenting a system or pro
echnique to estimate the average system load in a time-shared system. The solid lines indicate when each application is running.
gram changes what we are trying to measure. Obtaining more information, or obtaining higher resolution measurements, for instance, requires more instrumentation points in a program. However, more instrumentation causes there to be more perturbations in the program than there are in its uninstrumented execution behavior. These additional perturbations due to the additional instrumentation then make the data we collect less reliable. As a result, we are almost always forced to use insufficient data to infer the behavior of the system in which we are interested.
To further confound the situation, performance perturbations dueWholesale Ladies Cell Phoneinstrumentation are nonlinear and nonadditive. They are nonlinear in the sense that doubling the amount of instrumentation in a program will not necessarily double its impact on performance, for instance. Similarly, instrumentation perturbation is nonadditive in the sense that adding more instrumentation can cancel out the perturbation effects of other instrumentation. Or, in some situations, additional instrumentation can multiplicatively increase Wholesale Flip Cell Phoneperturbations.
For example, adding code to an application program to generate an instruction trace can significantly change the spatial and temporal patterns of its memory accesses. The trace-generation code will cause a large number of extra store instructions to be executed, for instance, which can cause the cache to be effectively flushed at each trace point. These frequent cache flushes will then increase the number of caches missed, which will substantially impact the overall performance. If additional instrumentation is added, however, it may be possible that the additional memory locations necessary for the instrumentation could change the pattern of conflict misses in the cache in such a way as to actually improve the cache performance perceived by the application. The bottom line is that Wholesale Bar Phoneeffects of adding instrumentation to a system being tested are entirely unpredictable.
Besides these direct changes to a program’s performance, instrumenting a program can cause more subtle indirect perturbations. For example,Wholesale TV Cell Phoneinstrumented program will take longer to execute than will the uninstrumented program. This increase in execution time will then cause it to experience more context switches than it would have experienced if it had not been instrumented. These additional context switches can substantially alter the program’s paging behavior, for instance, making the instrumented program behave substantially differently than the uninstrumented program.