A performance evaluation of a general parallel processing model

2009

Computer workloads have many attributes. When modeling these workloads it is often difficult to decide which attributes are important, and which can be abstracted away. In many cases, the modeler only includes attributes that are believed to be important, and ignores the rest. We argue, however, that this can lead to impaired workloads and unreliable system evaluations. Using parallel job scheduling as a case study, and daily cycles of activity as the attribute in dispute, we present two schedulers whose simulated performance seems identical without cycles, but then becomes significantly different when daily cycles are included in the workload. We trace this to the ability of one scheduler to prioritize interactive jobs, which leads to implicitly delaying less critical work to nighttime, when it can utilize resources that otherwise would have been left idle. Notably, this was not a design feature of this scheduler, but rather an emergent property that was not anticipated in advance. 100−299 >=300 © © © © 9 8 6 7 9 8 6 7 9 8 6 7 9 8 6 7 Figure 6: Characteristics of running and waiting jobs as sampled at different times of the day, with the CREASY and EASY schedulers (250 users). The situation with no daily cycles is shown on the right for comparison.

2010 Second International Conference on Computational Intelligence, Modelling and Simulation, 2010

From the queuing system approach, a compute intensive application can be defined as any application where the arrival rate of processes into the processors queue is greater than the overall departure rate of the processes from the processors. Such a compute intensive application is ideal for a parallel computer because there are always processes for the processors to execute. This paper, therefore, aims at using a novel and efficient queuing approach to model some of the performance metrics of the parallel processors, for compute intensive applications.

Proceedings of 6th Symposium on the Frontiers of Massively Parallel Computation (Frontiers '96), 1996

The analysis of workload traces from real production parallel machines can aid a wide variety of parallel processing research, providing a realistic basis for experimentation in the management of resources over an entire workload. We analyze a ve-month workload trace of an Intel Paragon machine supporting a production parallel workload at the San Diego Supercomputer Center (SDSC), comparing and contrasting it with a similar workload study of an Intel iPSC/860 machine at NASA Ames NAS. Our analysis of workload characteristics takes into account the job scheduling policies of the sites and focuses on characteristics such as job size distribution (job parallelism), resource usage, runtimes, submission patterns, and wait times. Despite fundamental di erences in the two machines and their respective usage environments, we observe a number of interesting similarities with respect to job size distribution, system utilization, runtime distribution, and interarrival time distribution. We hope to gain insight into the potential use of workload traces for evaluating resource management polices at supercomputing sites and for providing both real-world job streams and accurate stochastic workload models for use in simulation analysis of resource management policies.

Our goal in this section is to extend the previous performance analysis (based only on mean execution times) to include both fluctuations about mean values as well as ordering constraints or correlations between different steps and different jobs. This demands additional assumptions concerning the job arrival and execution time statistics, as well as the scheduling policy. We focus on a class of Markovian mathematical models where knowing only the present or current state of the system, not the entire history of the system operations, its future statistical behavior is determined. These are called Jackson queueing networks in honor of the pioneer researcher who first studied their behavior. If all possible system states and the associated state transition rates are exhaustively enumerated, oftentimes the number of states in such models explodes (into tens of trillions for straightforward applications such as a single terminal word processor system, and into far far greater numbers for more sophisticated applications such as airline reservation systems with hundreds of terminals, tens of communication lines, and upwards of one hundred disk spindles) to far exceed the storage and processing capabilities of the most powerful supercomputers (both available and envisioned!): it can be virtually impossible to extract any analytically tractable formulae or numerical approximations of different measures of performance. The only credible avenues for quantifying system performance for many people for years appeared to be simulation studies or building a prototype (which is another form of simulation, using the actual system but simulating the load). ----CHAPTER 6 JACKSON NETWORK ANALYSIS 3 Figure 6.1.System Block Diagram time of a job, denoted by E (T Q ).

Parallel Computing - Fundamentals and Applications - Proceedings of the International Conference ParCo99, 2000

This paper describes the collection and analysis of usage data for a large (hundreds of nodes) distributed memory machine over a period of 31 months during which 178,000 batch jobs were submitted. A number of data items were collected for each job, including the queue wait times, elapsed (wall clock) execution times, the number of nodes used, as well as the actual job CPU, system and wait times in node hours. This data set represents perhaps the most comprehensive such information on the use of a 100 Gflop parallel machine by a large (over 1,200 users in any given month) and diverse set of users. The results of this analysis provide some insights on how much machines are used and on workload profiles in terms of the number of nodes used, average queue wait times, elapsed and CPU times, as well as their distributions. A longitudinal analysis shows how these have changed over time and how scheduling policies affect user behavior. Some of these results confirm earlier studies, while others reveal new information. That knowledge has been used to develop a new scheduler for the system which has increased system node utilization from the 60% to the 90-95% range if there are sufficient jobs waiting in the queue.

Lecture Notes in Computer Science, 1995

Performance evaluation studies are to be an integral part of the design and tuning of parallel applications. Their structure and their behavior are the dominating factors. We propose a hierarchical approach to the systematic characterization of the workload of a parallel system, to be kept as modular and exible as possible. The methodology is based on three di erent, but related, layers: the application, the algorithm, and the routine layer. For each o f t h e s e l a yers di erent c haracteristics representing functional, sequential, parallel, and quantitative descriptions have been identi ed. Taking also architectural and mapping features into consideration, the hierarchical workload characterization can be used for any t ype of performance studies.

Job Scheduling Strategies for Parallel Processing, 2009

In parallel systems, similar jobs tend to arrive within bursty periods. This fact leads to the existence of the locality phenomenon, a persistent similarity between nearby jobs, in real parallel computer workloads. This important phenomenon deserves to be taken into account and used as a characteristic of any workload model. Regrettably, this property has received little if any attention of researchers and synthetic workloads used for performance evaluation to date often do not have locality. With respect to this research trend, Feitelson has suggested a general repetition approach to model locality in synthetic workloads . Using this approach, Li et al. recently introduced a new method for modeling temporal locality in workload attributes such as run time and memory . However, with the assumption that each job in the synthetic workload requires a single processor, the parallelism has not been taken into account in their study. In this paper, we propose a new model for parallel computer workloads based on their result. In our research, we firstly improve their model to control locality of a run time process better and then model the parallelism. The key idea for modeling the parallelism is to control the cross-correlation between the run time and the number of processors. Experimental results show that not only the cross-correlation is controlled well by our model, but also the marginal distribution can be fitted nicely. Furthermore, the locality feature is also obtained in our model.

IEEE Transactions on Computers, 1989

Lecture Notes in Computer Science, 1998

The evaluation of parallel job schedulers hinges on two things: the use of appropriate metrics, and the use of appropriate workloads on which the scheduler can operate. We argue that the focus should be on on-line open systems, and propose that a standard workload should be used as a benchmark for schedulers. This benchmark will specify distributions of parallelism and runtime, as found by analyzing accounting traces, and also internal structures that create different speedup and synchronization characteristics. As for metrics, we present some problems with slowdown and bounded slowdown that have been proposed recently.