A comprehensive model of the supercomputer workload

Abstract

As with any computer system, the performance of supercomputers depends upon the workloah that serve as their input. Unfortunately, however, there are many important aspects of the supercomputer workloads that have not been modeled, or that have being modeled only incipiently. This paper attacks this problem by considering requested time (and its relation with execution time) and the possibility of j o b cancellation, two aspects of the supercomputer workload that have not been modeled yet. Moreover, we also improve upon existing models for the arrival instant and partition size.

Workload Modeling for Computer Systems Performance Evaluation

“…An interesting compromise is to keep such data and explicitly include it in the workload model [137]. In other words, part of the workload model will be to model how often jobs fail or are aborted by the user who submitted them, or how often requests are made to download non-existent files.…”

“…A more general procedure that was advocated by Cirne and Berman is to use clustering as a means to distinguish between "normal" and "abnormal" data [137]. The methodology is to partition the workload log into days, and then to characterize each day by a vector of length n (specifically, this was applied to the modeling of the daily arrival cycle, and the vector contained the coefficients of a polynomial describing it).…”

“…Such models are nontrivial because parallel jobs have several distinct attributes, most notably their parallelism and runtime, which have to be modeled. Several models and benchmarks for this domain have been devised [739,230,369,187,137,453]. They complement the benchmarks of multithreaded applications on modern multicore systems that were discussed in Section 9.2.2.…”

“…Each job is then a rectangle in processors×time space: it uses a certain number of processors for a certain time. This enables the use of logs that contain just this data and the creation of models based on them [230,369,187,137,453]. Many of the examples in this book come from this domain.…”

Cloudy with a Chance of Load Spikes: Admission Control with Fuzzy Risk Assessments

“…In such scenarios, users have to provision the worst-case capacity as the upper bound. Third, users tend to overestimate their resource needs for the sake of safe execution, as illustrated by a supercomputing workload study that concluded that half of all jobs used less than 20% of the requested capacity [3].…”