Senior Solutions Architect – AI Infrastructure Networking

免费注册后解锁"你的匹配分 vs 高分画像"与个人技能缺口。

About the role We are seeking a Senior Solutions Architect with deep networking expertise to join the Voyager team in the Mirantis Office of the CTO. You will own the network solutioning of k0rdent AI, our platform for building and operating GPU clouds and AI factories, from metal to model. You will design how large GPU clusters are interconnected, how tenants are isolated, and how workloads reach the network from containers, virtual machines and bare metal nodes. You will turn that design into published reference architectures, working code and proofs of concept that run on real hardware with our customers and partners. As part of the Office of the CTO, research is a core part of the job. You will explore designs and technologies before customers ask for them, prototype ideas that may not ship, and challenge established practice when a better approach exists. Your work will shape where k0rdent AI goes next, not only how it is deployed today. This role combines research with hands-on, outward-facing work. You will split your time between exploring new designs, writing architecture, building and validating it in the lab, and explaining it to engineers, executives and conference audiences. You will work with Solutions Architects, Partner Management and Engineering colleagues across many countries and time zones. Key Responsibilities Reference architecture Design and publish network reference architectures and solution designs for k0rdent AI, covering data center topologies from a single rack to multi-thousand-GPU clusters. Define compute (east-west) fabrics on InfiniBand and on RoCEv2 Ethernet, including rail-optimised and fat-tree / Clos designs, oversubscription and failure-domain choices. Define front-end, storage and management networks, and how they connect to customer data centers, public clouds and hybrid environments. Specify multi-tenant isolation across the stack: InfiniBand partitions (PKeys), VRFs and EVPN-VXLAN on Ethernet, and Kubernetes-level network policy. Document scale limits and trade-offs (cost, performance, operability, vendor lock-in) with defensible reasoning a customer can follow. Host and platform networking Define how Linux hosts expose NICs, DPUs and SuperNICs to workloads: PCI passthrough, SR-IOV virtual functions, IOMMU groups, NUMA and GPU-NIC affinity. Design Kubernetes networking for AI workloads: primary and secondary CNIs, Multus, SR-IOV and RDMA device plugins, NVIDIA Network Operator, and Dynamic Resource Allocation (DRA). Design networking for virtual machines on KubeVirt, including passthrough and SR-IOV for GPU and RDMA traffic. Research and exploration Track and evaluate emerging technologies and standards in AI networking, such as Ultra Ethernet (UEC), SuperNIC and DPU offloads, scale-up and scale-across fabrics, and Kubernetes DRA network drivers. Prototype alternative designs in the lab, including ones customers have not yet asked for, and measure them against current practice. Challenge established or customer-preferred designs when evidence points to a better option, and make the case with data. Publish findings as internal research notes and design proposals and, where appropriate, as external papers, blog posts or talks. Code and proofs of concept Build and run proofs of concept with customers and partners, on Mirantis lab hardware and on customer sites, and report results against agreed success criteria. Write automation and tooling (Python, Go, Bash, Ansible, Helm, Kubernetes manifests, Terraform) that makes the reference architectures reproducible. Run validation and benchmarking of fabrics and host configurations (e.g. NCCL tests, perftest, ib_write_bw) and turn findings into design guidance. Customers, partners and community Act as the network subject matter expert in customer discovery, design reviews and architecture workshops, bringing thought leadership and new ideas rather than only reflecting current practice. Work with hardware and networking partners (e.g. NVIDIA, serv

关于该职位 我们正在寻找一位具备深厚网络专业知识的资深解决方案架构师，加入Mirantis CTO办公室的Voyager团队。你将负责k0rdent AI的网络解决方案设计，这是我们用于构建和运营GPU云及AI工厂的平台，覆盖从硬件到模型的全栈。你将设计大型GPU集群如何互联、租户如何隔离，以及工作负载如何从容器、虚拟机和裸金属节点接入网络。你将把这些设计转化为已发布的参考架构、可运行的代码和概念验证，并在客户和合作伙伴的真实硬件上运行。作为CTO办公室的一部分，研究是这项工作的核心内容。你将在客户提出需求之前探索设计和技术，对可能不会发布的想法进行原型验证，并在存在更好方法时挑战既定实践。你的工作将塑造k0rdent AI的未来方向，而不仅仅是它当前的部署方式。这个角色将研究与面向外部的实践工作相结合。你将在探索新设计、撰写架构文档、在实验室中构建和验证，以及向工程师、高管和会议听众进行讲解之间分配时间。你将与遍布多个国家和时区的解决方案架构师、合作伙伴管理和工程同事合作。 关键职责 参考架构 为k0rdent AI设计并发布网络参考架构和解决方案设计，覆盖从单机架到数千GPU集群的数据中心拓扑。 在InfiniBand和RoCEv2以太网上定义计算（东西向）网络架构，包括rail优化和fat-tree / Clos设计、超分比和故障域选择。 定义前端、存储和管理网络，以及它们如何连接到客户数据中心、公有云和混合环境。 指定跨栈的多租户隔离：InfiniBand分区（PKeys）、以太网上的VRF和EVPN-VXLAN，以及Kubernetes级别的网络策略。 记录规模限制和权衡（成本、性能、可运维性、供应商锁定），并提供客户可以理解的、有说服力的论证。 主机和平台网络 定义Linux主机如何向工作负载暴露NIC、DPU和SuperNIC：PCI直通、SR-IOV虚拟功能、IOMMU组、NUMA和GPU-NIC亲和性。 为AI工作负载设计Kubernetes网络：主CNI和辅助CNI、Multus、SR-IOV和RDMA设备插件、NVIDIA Network Operator，以及动态资源分配（DRA）。 为KubeVirt上的虚拟机设计网络，包括GPU和RDMA流量的直通和SR-IOV。 研究与探索 跟踪和评估AI网络中的新兴技术和标准，如Ultra Ethernet（UEC）、SuperNIC和DPU卸载、scale-up和scale-across网络架构，以及Kubernetes DRA网络驱动。 在实验室中对替代设计进行原型验证，包括客户尚未提出的设计，并将其与当前实践进行对比评估。 当证据指向更好的方案时，挑战既定或客户偏好的设计，并用数据说明理由。 将发现以内部研究笔记和设计提案的形式发布，并在适当情况下以外部论文、博客文章或演讲的形式发布。 代码和概念验证 与客户和合作伙伴一起，在Mirantis实验室硬件和客户现场构建并运行概念验证，并根据约定的成功标准报告结果。 编写自动化和工具（Python、Go、Bash、Ansible、Helm、Kubernetes清单、Terraform），使参考架构可复现。 对网络架构和主机配置进行验证和基准测试（例如NCCL测试、perftest、ib_write_bw），并将发现转化为设计指导。 客户、合作伙伴和社区 在客户需求探索、设计评审和架构研讨会中担任网络主题专家，带来思想领导力和新想法，而不仅仅是反映当前实践。 与硬件和网络合作伙伴（例如NVIDIA、serv

⚑ 岗位信息收录自公开招聘页（Himalayas），以雇主官网为准