PhD Position F/M Modelization of HPC Jobs and Resources to Minimize Energy Waste

Franck Corset (LJK, ASAR)

Soberness—in terms of electrical power—of Data Centers and High-Performance Computing (HPC) systems is becoming an important design issue, as the global energy consumption of Information Technologies is rising at consid- erable levels. Large-scale computing infrastructures are processing vaster amount of data or solving problems requiring vaster amount of computing power. The behavior of large scale infrastructures has become more variable and difficult to model, especially with respect to power consumption and application performance. Therefore, dealing with time variations and unpredictable disturbances demands to automate the management (i.e., configuration) of the infras- tructures. This automatic management can be done by periodically monitoring the state of the system, and updating the configuration to activate relevant mechanisms.

This work takes root in the field of autonomic computing [6], and aims at designing efficient feedback loops to automatically manage the resources of a HPC (high performance computing) infrastructure. The use of feedback loops is widespread in various fields of engineering, but recent in the computer science field.

Resource harvesting

The Resource and Job Management System (RJMS) is a key component to operate a HPC system [5, 7]. Users submit jobs: a description of their computation, data, and resource requirements. With respect to the resources status reported by the resource manager, the scheduler decides which resources to allocate to a job and assigns a time slot for the job’s execution.

The RJMS is however unable to fully exploit the resources in a HPC cluster: the presence of unused resources results from the limits of the scheduling of jobs. The allocation of resources to jobs, while respecting all constraints, leaves resources idle. Such inefficiency is sometimes referred as fragmentation. The loss of computing power resulting from the fragmentation represents an exploitable pool of resource.

CiGri [1] is a lightweight, scalable and fault-tolerant grid system that plugs into the RJMS. CiGri aims at minimizing the waste of computing resources to leverage the pool of unused resources. Yet, some computations (best-effort computations) can still lead to wasted resources if they are stopped before completion. In particular, by integrating information from the RJMS one could avoid to start best-effort computations if there is not enough time to execute them.

Modeling jobs and resources

The authors in [4] show that the jobs execution times are not necessary deterministic and can be modeled by an Exponential, Weibull, log-Normal or Normal distributions. Moreover, the high variance of execution times and the diversity of the jobs, whether in terms of the nature of the data, the application domain or the size of the problem have to be taking into account, by considering for instance a mixture of distributions (see [3]). Thus, we propose to apply some Machine Learning techniques, e.g., a clustering of all jobs submitted during the last decade, in order to take into account this heterogeneity. In a second step, we propose to take this information as a prior distribution in a Bayesian setting in order to improve the accuracy of the execution times estimations (see [8]). Furthermore, prior work mostly focuses on models that are job-centric and based on post-execution data [2]: they neglect to model the resources used for the computation.

In this work, we want to model the availability of the platform in order to insert the best-effort jobs in a frugal way. The design of models will have to consider that the allocation decisions are taken online, with partial information unveiling during the platform life-cycle.

References