Cherif Latreche
Open to platform, DevOps and SRE roles Rennes, France · remote-friendly

Platform / DevOps Engineer

I build and run Kubernetes and serverless platforms on bare metal, automate their deployment end to end, and make them observable.

Research Engineer at Inria, building E2Clab. PhD in Computer Science, 2026.

Explore experience Contact

Experience

Platforms I have built, operated and shipped.

  1. 2026 – Present
    Research Engineer, Edge-to-Cloud Platforms — Inria (Rennes, France)
    • Develop and maintain E2Clab, an open-source Python framework built on EnOSlib that deploys, runs and monitors distributed applications across edge and cloud testbeds.
    • Own the deployment automation for Grid'5000, Chameleon and FIT IoT-LAB: provisioning, configuration, network emulation and monitoring from a single declarative description.
    • Keep the project production-ready for external users through packaging, continuous integration, documentation and support.
  2. 2023 – 2026
    Platform Engineer (PhD), Cloud Infrastructure — INSA / IRISA (Rennes, France)
    • Operated bare-metal Kubernetes clusters and serverless platforms (Knative, Fission) as a continuously rebuilt environment; Python and EnOSlib automation provisioned up to 25 clusters from bare machines.
    • Designed a custom Kubernetes scheduler and a multi-cluster routing controller driven by live Prometheus metrics, reducing energy per request by up to 57% and p95 latency by up to 63% against the default scheduler.
    • Built observability and load testing end to end: Prometheus, Grafana and Kepler power metrics, open-loop traffic generators, latency and error reporting.
    • Taught hands-on Kubernetes and cloud computing labs to master's students.
  3. 2023
    Software Engineer — Institut Pasteur (Paris, France)
    • Delivered features and tests for clEsperanto, a GPU-accelerated image processing library.
    • Integrated and optimized OpenCL and CUDA kernels; built benchmarks to validate speedups and correctness.
  4. 2022
    Software Engineer — Namla (Orsay, France)
    • Implemented model-compression methods in TensorFlow to reduce model size and inference cost on constrained devices.
    • Quantified accuracy, latency and memory trade-offs and documented them to support deployment decisions.

Projects

Schedulers, controllers and tooling, each with a measured result.

E2Clab: edge-to-cloud experiment framework

  • Open-source Python framework built on EnOSlib for deploying, running and monitoring distributed applications across edge, fog and cloud resources.
  • A single declarative description drives provisioning, network emulation, monitoring and repeatable runs on Grid'5000, Chameleon and FIT IoT-LAB.
  • Supports scaling and parameter sweeps of a scenario, with provenance capture for every run.

Result. Provides a reproducible path from an application description to a running, monitored edge-to-cloud deployment.

Python · EnOSlib · Grid'5000 · Chameleon · FIT IoT-LAB
Code

Energy-aware Kubernetes scheduler

  • Custom Kubernetes scheduler written in Python, replacing the default scheduler for serverless functions on Fission.
  • Placement decisions rely on live CPU, memory and power metrics collected through Prometheus.
  • Benchmarked against the Kubernetes default scheduler and two alternative placement policies.

Result. 12% lower energy consumption and 7% lower latency than the Kubernetes default scheduler.

Python · Kubernetes · Fission · Prometheus
Code Paper · UCC 2024

Placement policy benchmark on Knative

  • Test bench on Knative and Kubernetes that replays identical traffic against five placement policies.
  • Four traffic patterns: steady, bursty, step and replayed Azure Functions traces.
  • Measures energy per request, p95 latency, failure rate and decision time for every policy.

Result. The best policy reduced energy per request by 27 to 57% and p95 latency by 37 to 63% over the reference policy.

Python · Kubernetes · Knative
Code Paper · ISPDC 2026

Multi-cluster request routing

  • Routing controller that selects the destination Kubernetes cluster for each function request.
  • Two designs evaluated: a central controller and one agent per cluster entry point.
  • Clusters run on Knative with Kepler energy metrics and emulated inter-site network delay.

Result. Both designs outperformed the strongest existing heuristic on energy and latency; the per-cluster agent decided in 36 ms against 86 ms for the central controller.

Python · Kubernetes · Knative · Kepler
Code Paper · CloudCom 2026

GPU-accelerated image processing benchmarks

  • Contributions to clEsperanto, a GPU-accelerated image processing library used in bio-imaging.
  • Integrated and optimized OpenCL and CUDA kernels for common image operations.
  • Designed benchmarks that measure speedups over CPU reference implementations and verify output correctness.

Result. Validated GPU speedups with correctness checks against CPU reference implementations on real image-processing workloads.

Python · C/C++ · OpenCL · CUDA
Code

About

I’m a platform and DevOps engineer. I run Kubernetes and serverless platforms on bare metal, automate their deployment end to end, and instrument them so decisions are driven by measured latency and cost. I completed a PhD in computer science in 2026 and now build E2Clab, an open-source framework for deploying and benchmarking distributed applications across edge and cloud infrastructure.

Tech stack

Containers & orchestration
KubernetesDocker
Serverless
KnativeFissionOpenFaaS
Infrastructure & automation
TerraformAnsibleEnOSlib
Cloud & testbeds
AWSGCPGrid'5000Chameleon
Observability
PrometheusGrafanaKepler
Programming
PythonGoC/C++Java
Machine Learning
TensorFlowStable-Baselines3Gymnasium

Education

2023 – 2026

PhD in Computer Science

INSA Rennes · IRISA

Thesis on energy-efficient serverless platforms on Kubernetes: scheduling, multi-cluster routing and performance evaluation on bare-metal clusters.

2022 – 2023

MSc in Computer Science

Université Paris-Saclay

Specialization in systems and software engineering, with coursework in distributed systems, GPU computing and machine learning.

Highlights

Awards, open source and peer-reviewed work.

Award2025

StartThèse Competition Winner

First place in a regional competition for research with industrial and entrepreneurial potential.

Pepite Bretagne
Award2024

International Mobility Grant

Funded a research stay and collaboration with a partner laboratory abroad.

Collège Doctoral de Bretagne
Open source2026

E2Clab maintainer

Develop and maintain an open-source framework for reproducible edge-to-cloud deployments.

Inria · GitLab
Publication2024

FoRLess: A Deep Reinforcement Learning-based Approach for FaaS Placement in Fog

Deep reinforcement learning places serverless functions across fog nodes to cut energy and latency.

IEEE/ACM UCC | Sharjah, UAE 2024
Publication2026

Do We Need Reinforcement Learning for Serverless Scheduling at the Edge? A Comparative Evaluation

Heuristic, bandit and deep RL schedulers compared on real edge clusters under four traffic patterns.

IEEE ISPDC | Hamburg, Germany 2026
Publication2026

Centralized vs. Decentralized Multi-Cluster FaaS Control using Reinforcement Learning

Central and per-cluster controllers route function requests across Kubernetes clusters.

IEEE CloudCom | Paris, France 2026

Let’s talk

Platform, DevOps and SRE roles, consulting, or anything Kubernetes.

Email GitHub LinkedIn Scholar

Rennes, France · CET