Performance modeling execution time

Resources are used to model assets or on a more general level to model every unit on which time consuming or quantity producing operations can be performed. Quants are assigned to resources and have executing times and costs on resources. A shift model and also lockup times can be assigned to a resource. 

Shift models and lockup times Every shift may have a performance factor f that changes an execution time T of a quant to T/f. The shift model may change after a few weeks or months. Therefore after this time the shift model can be simplified to the availability factor that expresses how much percent of the time can be used for production. It is possible to specify if a quant is interruptible by a shift pause or a lockup. This may depend on the product that is produced by the quant. Using the maximum allowed break for a quant is also a technique to model multiple single resources as one multiple resource. 

As an example, we outline a model of the performance of a Hartree-Fock (HE) algorithm in which the problem size is N, the number of basis functions. In other applications, the problem size could be the number of atoms, the number of particles in a simulation, and so on. The total execution time for the HE algorithm can be modeled as follows ... 

For data-parallel environments (SIMD programming model), performance can be represented using standard profiling techniques that associate fraction of overall execution time with particular pieces of code. As noted earlier, this programming model has limited utility for the applications we address, and we do not dwell on performance evaluation. 

We can approach performance modeling in different ways. One way is to formally derive the asymptotic behavior of the most time critical part of the program. The asymptotic behavior of an algorithm gives an estimate of the execution time as a function of problem size and of possibly other parameters. The notation that is commonly used is called big-oh [43]. For example, the statement that says that some method scales as means that there are positive constants c and... 

The DOMAR pubhsh-subscribe protocol, implemented in a set of SCORE forms, plays a central role in several important aspects of model building. Within the human performance models themselves, it is used to coordinate the concurrent operation of related procedures. It is frequently employed in the modeling of person-to-person communication, and it is the basis for the distributed execution, in real time or fast time, that DOMAR supports. Very large agent-based simulations may be developed in a distributed computing environment. 

We will use this algorithm instead of the all-to-all broadcast algorithm provided in the employed implementation of MPl because the latter algorithm displayed very irregular performance. A performance model for a matrix-vector multiplication that uses the all-to-all broadcast is discussed in section 6.4.1. The total execution time, the speedup, and the efficiency can then be expresssed as the following functions of p and n... 

In our performance models we will use a value of T = 0.99, fhat is, we will require that the probability that no process execution time exceeds that of the model equals 99%. Some care must be exercised in choosing an appropriate T value. A value too close to unity will cause Eqs. 7.5 and 7.6 to predict unrealistically high total execution times because there is no upper limit for the normal distribution. If T is too small, on fhe other hand, Eqs. 7.5 and 7.6 are likely to underestimate the total execution time. For simplicity, we will use an approximate form for fhe inverse error function, and for T = 0.99 we find the following expression fo be a very good approximafion to Eq. 7.10... 

Using the determined values for the various parameters, the performance models, Eqs. 9.10, 9.11, 9.13, and 9.14, yield the total execution times and speedups shown in Figures 9.4 and 9.5. The performance models predict PI... 

Keystroke-level performance model Models the performance (i.e., time) estimated for executing each of six defined actions, from which overall estimated task completion times can be computed Performance is modeled by assigning time values to each of six modeled actions... 

The failure time can be measured in terms of calendar time, execution time, or the number of failures per time interval (calendar or execution). Different models require different time data. Generally, it can be stated that using execution time increases the accuracy of the predictions, but requires a transformation into calendar time in order to be useful for some purposes. Planning of the test period is, for example, performed in terms of calendar time and not in exeeution time. Thus, there is a need for mapping between execution time and calendar time. Keeping track of actual test time. 

We note that our model consistently provides almost the same amount of performance degradation as the one observed in practice. This further supports the validity of our approach. An interesting consideration is that for smaller values of the key size the performance degradation is more dramatic. This is due to the fact that this key size affects the execution time of the Data Origin Authentication mechanism that is part of both configurations. Hence, while growing this size, the latter mechanism dominates in terms of execution time with respect to the Data Confidentiality, executed only in SC2, whose execution time does not vary with this key size. 

Common real-time models assume that tasks execute an infinite number of iterations of the same code (each iteration being called a job), with each release of the code being performed with specific time intervals (cyclic or periodic tasks), or triggered by an event (aperiodic or sporadic). It is also common that blocking (or voluntary-suspension) is not allowed inside the iteration of the loop. Blocking calls (such as delay-until or entry calls) are only allowed at each iteration start, to suspend the task until release. 

---