Reservation Strategies for Stochastic Jobs

Aupy, Guillaume; Gainaru, Ana; Honoré, Valentin; Raghavan, Padma; Robert, Yves; Sun, Hongyang

doi:10.1109/ipdps.2019.00027

Cited by 12 publications

(20 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For all the jobs where the first reservation was not large enough, the scheduler schedules the second reservation, and so on until all the jobs of the batch are executed. We have shown in a previous work [2] how to derive the optimal sequence of reservations to minimize the expected cost of a job, given the knowledge of the its execution profile (modeled by a probability distribution). Using a modified version of this algorithm that incorporates backfilling, we also showed [12] that the batch scheduler algorithm derived above was more efficient for both total system throughput, and average job response time!…”

Section: Motivationmentioning

confidence: 99%

“…In this work, we use the model from our previous work [2], [12]. We consider stochastic jobs, that is jobs whose execution time X is unknown but (i) deterministic: two successive execution on the same input have the same duration; and (ii) randomly and uniformly sampled from a probability distribution law D (PDF: f X ).…”

Section: A Stochastic Jobs and Reservation Strategymentioning

confidence: 99%

“…where the α part of this cost is the reservation cost, and the β part is the utilization cost. We have shown [2] that this cost function could indeed cover many different cost models (from execution time on an HPC platform to cost of usage in the cloud). The studies in this work are done for α = 1 and β = γ = 0.…”

Section: A Stochastic Jobs and Reservation Strategymentioning

confidence: 99%

“…In this work, we use α = 1, β = 0, γ = 0 (RESERVA-TIONONLY model [2]). In this case, the HPC system creates rigid reservations which are not adapted depending on how long the application is running (typical for several current large-scale systems).…”

Section: A Stochastic Jobs and Reservation Strategymentioning

confidence: 99%

“…We consider that we are given a sample data of size n: a) Discrete-Method: The most natural way to compute an estimate for the reservation strategy is to use the data as is: the n iid samples can be seen as a discrete distribution with n values, each of probability 1/n (assuming there is no redundant value, otherwise k/n if a value appears k times). We know that we can compute in polynomial time the optimal strategy for this distribution [2,Theorem 5]. We denote…”

Section: Estimators For Reservation Strategiesmentioning

confidence: 99%

See 4 more Smart Citations

Making Speculative Scheduling Robust to Incomplete Data

Gainaru

Pallez

2019

2019 IEEE/ACM 10th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Systems (ScalA)

Self Cite

View full text Add to dashboard Cite

In this work, we study the robustness of Speculative Scheduling to data incompleteness. Speculative scheduling has allowed to incorporate future types of applications into the design of HPC schedulers, specifically applications whose runtime is not perfectly known but can be modeled with probability distributions. Preliminary studies show the importance of speculative scheduling in dealing with stochastic applications when the application runtime model is completely known. In this work we show how one can extract enough information even from incomplete behavioral data for a given HPC applications so that speculative scheduling still performs well. Specifically, we show that for synthetic runtimes who follow usual probability distributions such as truncated normal or exponential, we can extract enough data from as little as 10 previous runs, to be within 5% of the solution which has exact information. For real traces of applications, the performance with 10 data points varies with the applications (within 20% of the full-knowledge solution), but converges fast (5% with 100 previous samples). Finally a side effect of this study is to show the importance of the theoretical results obtained on continuous probability distributions for speculative scheduling. Indeed, we observe that the solutions for such distributions are more robust to incomplete data than the solutions for discrete distributions.

show abstract

Section: Motivationmentioning

confidence: 99%

Section: A Stochastic Jobs and Reservation Strategymentioning

confidence: 99%

Section: A Stochastic Jobs and Reservation Strategymentioning

confidence: 99%

Section: A Stochastic Jobs and Reservation Strategymentioning

confidence: 99%

Section: Estimators For Reservation Strategiesmentioning

confidence: 99%

See 3 more Smart Citations

Making Speculative Scheduling Robust to Incomplete Data

Gainaru

Pallez

2019

2019 IEEE/ACM 10th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Systems (ScalA)

Self Cite

View full text Add to dashboard Cite

show abstract

Fine-Grained Scheduling for Containerized HPC Workloads in Kubernetes Clusters

Liu

Guitart

2022

2022 IEEE 24th Int Conf on High Performance Computing &Amp; Communications; 8th Int Conf on Data Science &Amp; Systems; 20th In

View full text Add to dashboard Cite

Containerization technology offers lightweight OSlevel virtualization, and enables portability, reproducibility, and flexibility by packing applications with low performance overhead and low effort to maintain and scale them. Moreover, container orchestrators (e.g., Kubernetes) are widely used in the Cloud to manage large clusters running many containerized applications. However, scheduling policies that consider the performance nuances of containerized High Performance Computing (HPC) workloads have not been well-explored yet. This paper conducts fine-grained scheduling policies for containerized HPC workloads in Kubernetes clusters, focusing especially on partitioning each job into a suitable multi-container deployment according to the application profile. We implement our scheduling schemes on different layers of management (application and infrastructure), so that each component has its own focus and algorithms but still collaborates with others. Our results show that our fine-grained scheduling policies outperform baseline and baseline with CPU/memory affinity enabled policies, reducing the overall response time by 35% and 19%, respectively, and also improving the makespan by 34% and 11%, respectively. They also provide better usability and flexibility to specify HPC workloads than other comparable HPC Cloud frameworks, while providing better scheduling efficiency thanks to their multi-layered approach.

show abstract